+73 -11

Cargo.lock

··· 152 152 checksum = "2261d10cca569e4643e526d8dc2e62e433cc8aba21ab764233731f8d369bf394" 153 153 154 154 [[package]] 155 155 + name = "block-buffer" 156 156 + version = "0.10.4" 157 157 + source = "registry+https://github.com/rust-lang/crates.io-index" 158 158 + checksum = "3078c7629b62d3f0439517fa394996acacc5cbc91c5a20d8c658e77abd503a71" 159 159 + dependencies = [ 160 160 + "generic-array", 161 161 + ] 162 162 + 163 163 + [[package]] 155 164 name = "bumpalo" 156 165 version = "3.19.0" 157 166 source = "registry+https://github.com/rust-lang/crates.io-index" ··· 287 296 ] 288 297 289 298 [[package]] 299 299 + name = "cpufeatures" 300 300 + version = "0.2.17" 301 301 + source = "registry+https://github.com/rust-lang/crates.io-index" 302 302 + checksum = "59ed5838eebb26a2bb2e58f6d5b5316989ae9d08bab10e0e6d103e656d1b0280" 303 303 + dependencies = [ 304 304 + "libc", 305 305 + ] 306 306 + 307 307 + [[package]] 290 308 name = "criterion" 291 309 version = "0.7.0" 292 310 source = "registry+https://github.com/rust-lang/crates.io-index" ··· 352 370 checksum = "460fbee9c2c2f33933d720630a6a0bac33ba7053db5344fac858d4b8952d77d5" 353 371 354 372 [[package]] 373 373 + name = "crypto-common" 374 374 + version = "0.1.6" 375 375 + source = "registry+https://github.com/rust-lang/crates.io-index" 376 376 + checksum = "1bfb12502f3fc46cca1bb51ac28df9d618d813cdc3d2f25b9fe775a34af26bb3" 377 377 + dependencies = [ 378 378 + "generic-array", 379 379 + "typenum", 380 380 + ] 381 381 + 382 382 + [[package]] 355 383 name = "data-encoding" 356 384 version = "2.9.0" 357 385 source = "registry+https://github.com/rust-lang/crates.io-index" ··· 375 403 dependencies = [ 376 404 "data-encoding", 377 405 "syn 2.0.106", 406 406 + ] 407 407 + 408 408 + [[package]] 409 409 + name = "digest" 410 410 + version = "0.10.7" 411 411 + source = "registry+https://github.com/rust-lang/crates.io-index" 412 412 + checksum = "9ed9a281f7bc9b7576e61468ba615a66a5c8cfdff42420a70aa82701a3b1e292" 413 413 + dependencies = [ 414 414 + "block-buffer", 415 415 + "crypto-common", 378 416 ] 379 417 380 418 [[package]] ··· 530 568 ] 531 569 532 570 [[package]] 571 571 + name = "generic-array" 572 572 + version = "0.14.9" 573 573 + source = "registry+https://github.com/rust-lang/crates.io-index" 574 574 + checksum = "4bb6743198531e02858aeaea5398fcc883e71851fcbcb5a2f773e2fb6cb1edf2" 575 575 + dependencies = [ 576 576 + "typenum", 577 577 + "version_check", 578 578 + ] 579 579 + 580 580 + [[package]] 533 581 name = "getrandom" 534 582 version = "0.3.3" 535 583 source = "registry+https://github.com/rust-lang/crates.io-index" ··· 937 985 ] 938 986 939 987 [[package]] 940 940 - name = "redb" 941 941 - version = "3.1.0" 942 942 - source = "registry+https://github.com/rust-lang/crates.io-index" 943 943 - checksum = "ae323eb086579a3769daa2c753bb96deb95993c534711e0dbe881b5192906a06" 944 944 - dependencies = [ 945 945 - "libc", 946 946 - ] 947 947 - 948 948 - [[package]] 949 988 name = "redox_syscall" 950 989 version = "0.5.18" 951 990 source = "registry+https://github.com/rust-lang/crates.io-index" ··· 985 1024 986 1025 [[package]] 987 1026 name = "repo-stream" 988 988 - version = "0.1.1" 1027 1027 + version = "0.2.2" 989 1028 dependencies = [ 990 1029 "bincode", 991 1030 "clap", ··· 997 1036 "iroh-car", 998 1037 "log", 999 1038 "multibase", 1000 1000 - "redb", 1001 1039 "rusqlite", 1002 1040 "serde", 1003 1041 "serde_bytes", 1004 1042 "serde_ipld_dagcbor", 1043 1043 + "sha2", 1005 1044 "tempfile", 1006 1045 "thiserror 2.0.17", 1007 1046 "tokio", ··· 1133 1172 ] 1134 1173 1135 1174 [[package]] 1175 1175 + name = "sha2" 1176 1176 + version = "0.10.9" 1177 1177 + source = "registry+https://github.com/rust-lang/crates.io-index" 1178 1178 + checksum = "a7507d819769d01a365ab707794a4084392c824f54a7a6a7862f8c3d0892b283" 1179 1179 + dependencies = [ 1180 1180 + "cfg-if", 1181 1181 + "cpufeatures", 1182 1182 + "digest", 1183 1183 + ] 1184 1184 + 1185 1185 + [[package]] 1136 1186 name = "signal-hook-registry" 1137 1187 version = "1.4.6" 1138 1188 source = "registry+https://github.com/rust-lang/crates.io-index" ··· 1286 1336 ] 1287 1337 1288 1338 [[package]] 1339 1339 + name = "typenum" 1340 1340 + version = "1.19.0" 1341 1341 + source = "registry+https://github.com/rust-lang/crates.io-index" 1342 1342 + checksum = "562d481066bde0658276a35467c4af00bdc6ee726305698a55b86e61d7ad82bb" 1343 1343 + 1344 1344 + [[package]] 1289 1345 name = "unicode-ident" 1290 1346 version = "1.0.19" 1291 1347 source = "registry+https://github.com/rust-lang/crates.io-index" ··· 1320 1376 version = "0.2.15" 1321 1377 source = "registry+https://github.com/rust-lang/crates.io-index" 1322 1378 checksum = "accd4ea62f7bb7a82fe23066fb0957d48ef677f6eeb8215f372f52e48bb32426" 1379 1379 + 1380 1380 + [[package]] 1381 1381 + name = "version_check" 1382 1382 + version = "0.9.5" 1383 1383 + source = "registry+https://github.com/rust-lang/crates.io-index" 1384 1384 + checksum = "0b928f33d975fc6ad9f86c8f283853ad26bdd5b10b7f1542aa2fa15e2289105a" 1323 1385 1324 1386 [[package]] 1325 1387 name = "virtue"

+7 -4

Cargo.toml

··· 1 1 [package] 2 2 name = "repo-stream" 3 3 - version = "0.1.1" 3 3 + version = "0.2.2" 4 4 edition = "2024" 5 5 license = "MIT OR Apache-2.0" 6 6 - description = "Fast and robust atproto CAR file processing in rust" 6 6 + description = "A robust CAR file -> MST walker for atproto" 7 7 repository = "https://tangled.org/@microcosm.blue/repo-stream" 8 8 9 9 [dependencies] ··· 14 14 iroh-car = "0.5.1" 15 15 log = "0.4.28" 16 16 multibase = "0.9.2" 17 17 - redb = "3.1.0" 18 17 rusqlite = "0.37.0" 19 18 serde = { version = "1.0.228", features = ["derive"] } 20 19 serde_bytes = "0.11.19" 21 20 serde_ipld_dagcbor = "0.6.4" 21 21 + sha2 = "0.10.9" 22 22 thiserror = "2.0.17" 23 23 - tokio = { version = "1.47.1", features = ["rt"] } 23 23 + tokio = { version = "1.47.1", features = ["rt", "sync"] } 24 24 25 25 [dev-dependencies] 26 26 clap = { version = "4.5.48", features = ["derive"] } ··· 33 33 [profile.profiling] 34 34 inherits = "release" 35 35 debug = true 36 36 + 37 37 + # [profile.release] 38 38 + # debug = true 36 39 37 40 [[bench]] 38 41 name = "non-huge-cars"

+12 -21

benches/huge-car.rs

··· 1 1 extern crate repo_stream; 2 2 - use futures::TryStreamExt; 3 3 - use iroh_car::CarReader; 4 4 - use std::convert::Infallible; 2 2 + use repo_stream::Driver; 5 3 use std::path::{Path, PathBuf}; 6 4 7 5 use criterion::{Criterion, criterion_group, criterion_main}; ··· 20 18 }); 21 19 } 22 20 23 23 - async fn drive_car(filename: impl AsRef<Path>) { 21 21 + async fn drive_car(filename: impl AsRef<Path>) -> usize { 24 22 let reader = tokio::fs::File::open(filename).await.unwrap(); 25 23 let reader = tokio::io::BufReader::new(reader); 26 26 - let reader = CarReader::new(reader).await.unwrap(); 27 24 28 28 - let root = reader 29 29 - .header() 30 30 - .roots() 31 31 - .first() 32 32 - .ok_or("missing root") 25 25 + let mut driver = match Driver::load_car(reader, |block| block.len(), 1024) 26 26 + .await 33 27 .unwrap() 34 34 - .clone(); 35 35 - 36 36 - let stream = std::pin::pin!(reader.stream()); 37 37 - 38 38 - let (_commit, v) = 39 39 - repo_stream::drive::Vehicle::init(root, stream, |block| Ok::<_, Infallible>(block.len())) 40 40 - .await 41 41 - .unwrap(); 42 42 - let mut record_stream = std::pin::pin!(v.stream()); 28 28 + { 29 29 + Driver::Memory(_, mem_driver) => mem_driver, 30 30 + Driver::Disk(_) => panic!("not doing disk for benchmark"), 31 31 + }; 43 32 44 44 - while let Some(_) = record_stream.try_next().await.unwrap() { 45 45 - // just here for the drive 33 33 + let mut n = 0; 34 34 + while let Some(pairs) = driver.next_chunk(256).await.unwrap() { 35 35 + n += pairs.len(); 46 36 } 37 37 + n 47 38 } 48 39 49 40 criterion_group!(benches, criterion_benchmark);

+16 -22

benches/non-huge-cars.rs

··· 1 1 extern crate repo_stream; 2 2 - use futures::TryStreamExt; 3 3 - use iroh_car::CarReader; 4 4 - use std::convert::Infallible; 2 2 + use repo_stream::Driver; 5 3 6 4 use criterion::{Criterion, criterion_group, criterion_main}; 7 5 6 6 + const EMPTY_CAR: &'static [u8] = include_bytes!("../car-samples/empty.car"); 8 7 const TINY_CAR: &'static [u8] = include_bytes!("../car-samples/tiny.car"); 9 8 const LITTLE_CAR: &'static [u8] = include_bytes!("../car-samples/little.car"); 10 9 const MIDSIZE_CAR: &'static [u8] = include_bytes!("../car-samples/midsize.car"); ··· 15 14 .build() 16 15 .expect("Creating runtime failed"); 17 16 17 17 + c.bench_function("empty-car", |b| { 18 18 + b.to_async(&rt).iter(async || drive_car(EMPTY_CAR).await) 19 19 + }); 18 20 c.bench_function("tiny-car", |b| { 19 21 b.to_async(&rt).iter(async || drive_car(TINY_CAR).await) 20 22 }); ··· 26 28 }); 27 29 } 28 30 29 29 - async fn drive_car(bytes: &[u8]) { 30 30 - let reader = CarReader::new(bytes).await.unwrap(); 31 31 - 32 32 - let root = reader 33 33 - .header() 34 34 - .roots() 35 35 - .first() 36 36 - .ok_or("missing root") 31 31 + async fn drive_car(bytes: &[u8]) -> usize { 32 32 + let mut driver = match Driver::load_car(bytes, |block| block.len(), 32) 33 33 + .await 37 34 .unwrap() 38 38 - .clone(); 39 39 - 40 40 - let stream = std::pin::pin!(reader.stream()); 35 35 + { 36 36 + Driver::Memory(_, mem_driver) => mem_driver, 37 37 + Driver::Disk(_) => panic!("not benching big cars here"), 38 38 + }; 41 39 42 42 - let (_commit, v) = 43 43 - repo_stream::drive::Vehicle::init(root, stream, |block| Ok::<_, Infallible>(block.len())) 44 44 - .await 45 45 - .unwrap(); 46 46 - let mut record_stream = std::pin::pin!(v.stream()); 47 47 - 48 48 - while let Some(_) = record_stream.try_next().await.unwrap() { 49 49 - // just here for the drive 40 40 + let mut n = 0; 41 41 + while let Some(pairs) = driver.next_chunk(256).await.unwrap() { 42 42 + n += pairs.len(); 50 43 } 44 44 + n 51 45 } 52 46 53 47 criterion_group!(benches, criterion_benchmark);

car-samples/empty.car

This is a binary file and will not be displayed.

+66 -29

examples/disk-read-file/main.rs

··· 1 1 + /*! 2 2 + Read a CAR file by spilling to disk 3 3 + */ 4 4 + 1 5 extern crate repo_stream; 2 6 use clap::Parser; 3 3 - use repo_stream::disk::RedbStore; 4 4 - use repo_stream::drive::Processable; 5 5 - use serde::{Deserialize, Serialize}; 7 7 + use repo_stream::{DiskBuilder, Driver, DriverBuilder}; 6 8 use std::path::PathBuf; 7 7 - 8 8 - type Result<T> = std::result::Result<T, Box<dyn std::error::Error>>; 9 9 + use std::time::Instant; 9 10 10 11 #[derive(Debug, Parser)] 11 12 struct Args { ··· 15 16 tmpfile: PathBuf, 16 17 } 17 18 18 18 - #[derive(Clone, Serialize, Deserialize)] 19 19 - struct S(usize); 20 20 - 21 21 - impl Processable for S {} 22 22 - 23 19 #[tokio::main] 24 24 - async fn main() -> Result<()> { 20 20 + async fn main() -> Result<(), Box<dyn std::error::Error>> { 25 21 env_logger::init(); 26 22 27 23 let Args { car, tmpfile } = Args::parse(); 24 24 + 25 25 + // repo-stream takes an AsyncRead as input. wrapping a filesystem read in 26 26 + // BufReader can provide a really significant performance win. 28 27 let reader = tokio::fs::File::open(car).await?; 29 28 let reader = tokio::io::BufReader::new(reader); 30 29 31 31 - let mut driver = 32 32 - match repo_stream::drive::load_car(reader, |block| S(block.len()), 1024).await? { 33 33 - repo_stream::drive::Vehicle::Lil(_, _) => panic!("try this on a bigger car"), 34 34 - repo_stream::drive::Vehicle::Big(big_stuff) => { 35 35 - let disk_store = RedbStore::new(tmpfile); 36 36 - let (commit, driver) = big_stuff.finish_loading(disk_store).await?; 37 37 - log::warn!("big: {:?}", commit); 38 38 - driver 39 39 - } 40 40 - }; 30 30 + log::info!("hello! reading the car..."); 31 31 + let t0 = Instant::now(); 32 32 + 33 33 + // in this example we only bother handling CARs that are too big for memory 34 34 + // `noop` helper means: do no block processing, store the raw blocks 35 35 + let driver = match DriverBuilder::new() 36 36 + .with_mem_limit_mb(10) // how much memory can be used before disk spill 37 37 + .load_car(reader) 38 38 + .await? 39 39 + { 40 40 + Driver::Memory(_, _) => panic!("try this on a bigger car"), 41 41 + Driver::Disk(big_stuff) => { 42 42 + // we reach here if the repo was too big and needs to be spilled to 43 43 + // disk to continue 44 44 + 45 45 + // set up a disk store we can spill to 46 46 + let disk_store = DiskBuilder::new().open(tmpfile).await?; 47 47 + 48 48 + // do the spilling, get back a (similar) driver 49 49 + let (commit, driver) = big_stuff.finish_loading(disk_store).await?; 50 50 + 51 51 + // at this point you might want to fetch the account's signing key 52 52 + // via the DID from the commit, and then verify the signature. 53 53 + log::warn!("big's comit ({:?}): {:?}", t0.elapsed(), commit); 41 54 42 42 - println!("hello!"); 55 55 + // pop the driver back out to get some code indentation relief 56 56 + driver 57 57 + } 58 58 + }; 43 59 60 60 + // collect some random stats about the blocks 44 61 let mut n = 0; 45 45 - loop { 46 46 - let (d, Some(pairs)) = driver.next_chunk(256).await? else { 47 47 - break; 48 48 - }; 49 49 - driver = d; 62 62 + let mut zeros = 0; 63 63 + 64 64 + log::info!("walking..."); 65 65 + 66 66 + // this example uses the disk driver's channel mode: the tree walking is 67 67 + // spawned onto a blocking thread, and we get chunks of rkey+blocks back 68 68 + let (mut rx, join) = driver.to_channel(512); 69 69 + while let Some(r) = rx.recv().await { 70 70 + let pairs = r?; 71 71 + 72 72 + // keep a count of the total number of blocks seen 50 73 n += pairs.len(); 51 51 - // log::info!("got {rkey:?}"); 74 74 + 75 75 + for (_, block) in pairs { 76 76 + // for each block, count how many bytes are equal to '0' 77 77 + // (this is just an example, you probably want to do something more 78 78 + // interesting) 79 79 + zeros += block.into_iter().filter(|&b| b == b'0').count() 80 80 + } 52 81 } 53 53 - log::info!("bye! {n}"); 82 82 + 83 83 + log::info!("arrived! ({:?}) joining rx...", t0.elapsed()); 84 84 + 85 85 + // clean up the database. would be nice to do this in drop so it happens 86 86 + // automatically, but some blocking work happens, so that's not allowed in 87 87 + // async rust. 🤷‍♀️ 88 88 + join.await?.reset_store().await?; 89 89 + 90 90 + log::info!("done. n={n} zeros={zeros}"); 54 91 55 92 Ok(()) 56 93 }

+14 -6

examples/read-file/main.rs

··· 1 1 + /*! 2 2 + Read a CAR file with in-memory processing 3 3 + */ 4 4 + 1 5 extern crate repo_stream; 2 6 use clap::Parser; 7 7 + use repo_stream::{Driver, DriverBuilder}; 3 8 use std::path::PathBuf; 4 9 5 10 type Result<T> = std::result::Result<T, Box<dyn std::error::Error>>; ··· 18 23 let reader = tokio::fs::File::open(file).await?; 19 24 let reader = tokio::io::BufReader::new(reader); 20 25 21 21 - let (commit, mut driver) = 22 22 - match repo_stream::drive::load_car(reader, |block| block.len(), 1024 * 1024).await? { 23 23 - repo_stream::drive::Vehicle::Lil(commit, mem_driver) => (commit, mem_driver), 24 24 - repo_stream::drive::Vehicle::Big(_) => panic!("can't handle big cars yet"), 25 25 - }; 26 26 + let (commit, mut driver) = match DriverBuilder::new() 27 27 + .with_block_processor(|block| block.len()) 28 28 + .load_car(reader) 29 29 + .await? 30 30 + { 31 31 + Driver::Memory(commit, mem_driver) => (commit, mem_driver), 32 32 + Driver::Disk(_) => panic!("this example doesn't handle big CARs"), 33 33 + }; 26 34 27 35 log::info!("got commit: {commit:?}"); 28 36 ··· 31 39 n += pairs.len(); 32 40 // log::info!("got {rkey:?}"); 33 41 } 34 34 - log::info!("bye! {n}"); 42 42 + log::info!("bye! total records={n}"); 35 43 36 44 Ok(()) 37 45 }

+70 -2

readme.md

··· 1 1 # repo-stream 2 2 3 3 - Fast and (aspirationally) robust atproto CAR file processing in rust 3 3 + A robust CAR file -> MST walker for atproto 4 4 + 5 5 + [![Crates.io][crates-badge]](https://crates.io/crates/repo-stream) 6 6 + [![Documentation][docs-badge]](https://docs.rs/repo-stream) 7 7 + [![Sponsor][sponsor-badge]](https://github.com/sponsors/uniphil) 8 8 + 9 9 + [crates-badge]: https://img.shields.io/crates/v/repo-stream.svg 10 10 + [docs-badge]: https://docs.rs/repo-stream/badge.svg 11 11 + [sponsor-badge]: https://img.shields.io/badge/at-microcosm-b820f9?labelColor=b820f9&logo=githubsponsors&logoColor=fff 12 12 + 13 13 + ```rust 14 14 + use repo_stream::{Driver, DriverBuilder, DriveError, DiskBuilder}; 15 15 + 16 16 + #[tokio::main] 17 17 + async fn main() -> Result<(), DriveError> { 18 18 + // repo-stream takes any AsyncRead as input, like a tokio::fs::File 19 19 + let reader = tokio::fs::File::open("repo.car".into()).await?; 20 20 + let reader = tokio::io::BufReader::new(reader); 21 21 + 22 22 + // example repo workload is simply counting the total record bytes 23 23 + let mut total_size = 0; 24 24 + 25 25 + match DriverBuilder::new() 26 26 + .with_mem_limit_mb(10) 27 27 + .with_block_processor(|rec| rec.len()) // block processing: just extract the raw record size 28 28 + .load_car(reader) 29 29 + .await? 30 30 + { 31 31 + 32 32 + // if all blocks fit within memory 33 33 + Driver::Memory(_commit, mut driver) => { 34 34 + while let Some(chunk) = driver.next_chunk(256).await? { 35 35 + for (_rkey, size) in chunk { 36 36 + total_size += size; 37 37 + } 38 38 + } 39 39 + }, 40 40 + 41 41 + // if the CAR was too big for in-memory processing 42 42 + Driver::Disk(paused) => { 43 43 + // set up a disk store we can spill to 44 44 + let store = DiskBuilder::new().open("some/path.db".into()).await?; 45 45 + // do the spilling, get back a (similar) driver 46 46 + let (_commit, mut driver) = paused.finish_loading(store).await?; 47 47 + 48 48 + while let Some(chunk) = driver.next_chunk(256).await? { 49 49 + for (_rkey, size) in chunk { 50 50 + total_size += size; 51 51 + } 52 52 + } 53 53 + 54 54 + // clean up the disk store (drop tables etc) 55 55 + driver.reset_store().await?; 56 56 + } 57 57 + }; 58 58 + println!("sum of size of all records: {total_size}"); 59 59 + Ok(()) 60 60 + } 61 61 + ``` 62 62 + 63 63 + more recent todo 64 64 + 65 65 + - [ ] get an *emtpy* car for the test suite 66 66 + - [x] implement a max size on disk limit 67 67 + 68 68 + 69 69 + ----- 70 70 + 71 71 + older stuff (to clean up): 4 72 5 73 6 74 current car processing times (records processed into their length usize, phil's dev machine): ··· 27 95 -> yeah the commit is returned from init 28 96 - [ ] spec compliance todos 29 97 - [x] assert that keys are ordered and fail if not 30 30 - - [ ] verify node mst depth from key (possibly pending [interop test fixes](https://github.com/bluesky-social/atproto-interop-tests/issues/5)) 98 98 + - [x] verify node mst depth from key (possibly pending [interop test fixes](https://github.com/bluesky-social/atproto-interop-tests/issues/5)) 31 99 - [ ] performance todos 32 100 - [x] consume the serialized nodes into a mutable efficient format 33 101 - [ ] maybe customize the deserialize impl to do that directly?

+175 -160

src/disk.rs

··· 1 1 - use redb::ReadableDatabase; 2 2 - use rusqlite::OptionalExtension; 3 3 - use std::error::Error; 4 4 - use std::path::PathBuf; 1 1 + /*! 2 2 + Disk storage for blocks on disk 5 3 6 6 - pub trait StorageErrorBase: Error + Send + 'static {} 7 7 - 8 8 - /// high level potential storage resource 9 9 - /// 10 10 - /// separating this allows (hopefully) implementing a storage pool that can 11 11 - /// async-block when until a member is available to use 12 12 - pub trait DiskStore { 13 13 - type StorageError: StorageErrorBase + Send; 14 14 - type Access: DiskAccess<StorageError = Self::StorageError>; 15 15 - fn get_access(&mut self) -> impl Future<Output = Result<Self::Access, Self::StorageError>>; 16 16 - } 17 17 - 18 18 - /// actual concrete access to disk storage 19 19 - pub trait DiskAccess: Send { 20 20 - type StorageError: StorageErrorBase; 4 4 + Currently this uses sqlite. In testing sqlite wasn't the fastest, but it seemed 5 5 + to be the best behaved in terms of both on-disk space usage and memory usage. 21 6 22 22 - fn get_writer(&mut self) -> Result<impl DiskWriter<Self::StorageError>, Self::StorageError>; 7 7 + ```no_run 8 8 + # use repo_stream::{DiskBuilder, DiskError}; 9 9 + # #[tokio::main] 10 10 + # async fn main() -> Result<(), DiskError> { 11 11 + let store = DiskBuilder::new() 12 12 + .with_cache_size_mb(32) 13 13 + .with_max_stored_mb(1024) // errors when >1GiB of processed blocks are inserted 14 14 + .open("/some/path.db".into()).await?; 15 15 + # Ok(()) 16 16 + # } 17 17 + ``` 18 18 + */ 23 19 24 24 - fn get_reader( 25 25 - &self, 26 26 - ) -> Result<impl DiskReader<StorageError = Self::StorageError>, Self::StorageError>; 20 20 + use crate::drive::DriveError; 21 21 + use rusqlite::OptionalExtension; 22 22 + use std::path::PathBuf; 27 23 28 28 - // TODO: force a cleanup implementation? 24 24 + #[derive(Debug, thiserror::Error)] 25 25 + pub enum DiskError { 26 26 + /// A wrapped database error 27 27 + /// 28 28 + /// (The wrapped err should probably be obscured to remove public-facing 29 29 + /// sqlite bits) 30 30 + #[error(transparent)] 31 31 + DbError(#[from] rusqlite::Error), 32 32 + /// A tokio blocking task failed to join 33 33 + #[error("Failed to join a tokio blocking task: {0}")] 34 34 + JoinError(#[from] tokio::task::JoinError), 35 35 + /// The total size of stored blocks exceeded the allowed size 36 36 + /// 37 37 + /// If you need to process *really* big CARs, you can configure a higher 38 38 + /// limit. 39 39 + #[error("Maximum disk size reached")] 40 40 + MaxSizeExceeded, 41 41 + #[error("this error was replaced, seeing this is a bug.")] 42 42 + #[doc(hidden)] 43 43 + Stolen, 29 44 } 30 45 31 31 - pub trait DiskWriter<E: StorageErrorBase> { 32 32 - fn put(&mut self, key: Vec<u8>, val: Vec<u8>) -> Result<(), E>; 46 46 + impl DiskError { 47 47 + /// hack for ownership challenges with the disk driver 48 48 + pub(crate) fn steal(&mut self) -> Self { 49 49 + let mut swapped = DiskError::Stolen; 50 50 + std::mem::swap(self, &mut swapped); 51 51 + swapped 52 52 + } 33 53 } 34 54 35 35 - pub trait DiskReader { 36 36 - type StorageError: StorageErrorBase; 37 37 - fn get(&mut self, key: Vec<u8>) -> Result<Option<Vec<u8>>, Self::StorageError>; 55 55 + /// Builder-style disk store setup 56 56 + #[derive(Debug, Clone)] 57 57 + pub struct DiskBuilder { 58 58 + /// Database in-memory cache allowance 59 59 + /// 60 60 + /// Default: 32 MiB 61 61 + pub cache_size_mb: usize, 62 62 + /// Database stored block size limit 63 63 + /// 64 64 + /// Default: 10 GiB 65 65 + /// 66 66 + /// Note: actual size on disk may be more, but should approximately scale 67 67 + /// with this limit 68 68 + pub max_stored_mb: usize, 38 69 } 39 70 40 40 - ///////////////// sqlite 41 41 - 42 42 - pub struct SqliteStore { 43 43 - path: PathBuf, 71 71 + impl Default for DiskBuilder { 72 72 + fn default() -> Self { 73 73 + Self { 74 74 + cache_size_mb: 32, 75 75 + max_stored_mb: 10 * 1024, // 10 GiB 76 76 + } 77 77 + } 44 78 } 45 79 46 46 - impl SqliteStore { 47 47 - pub fn new(path: PathBuf) -> Self { 48 48 - Self { path } 80 80 + impl DiskBuilder { 81 81 + /// Begin configuring the storage with defaults 82 82 + pub fn new() -> Self { 83 83 + Default::default() 84 84 + } 85 85 + /// Set the in-memory cache allowance for the database 86 86 + /// 87 87 + /// Default: 32 MiB 88 88 + pub fn with_cache_size_mb(mut self, size: usize) -> Self { 89 89 + self.cache_size_mb = size; 90 90 + self 91 91 + } 92 92 + /// Set the approximate stored block size limit 93 93 + /// 94 94 + /// Default: 10 GiB 95 95 + pub fn with_max_stored_mb(mut self, max: usize) -> Self { 96 96 + self.max_stored_mb = max; 97 97 + self 98 98 + } 99 99 + /// Open and initialize the actual disk storage 100 100 + pub async fn open(&self, path: PathBuf) -> Result<DiskStore, DiskError> { 101 101 + DiskStore::new(path, self.cache_size_mb, self.max_stored_mb).await 49 102 } 50 103 } 51 104 52 52 - impl StorageErrorBase for rusqlite::Error {} 105 105 + /// On-disk block storage 106 106 + pub struct DiskStore { 107 107 + conn: rusqlite::Connection, 108 108 + max_stored: usize, 109 109 + stored: usize, 110 110 + } 53 111 54 54 - impl DiskStore for SqliteStore { 55 55 - type StorageError = rusqlite::Error; 56 56 - type Access = SqliteAccess; 57 57 - async fn get_access(&mut self) -> Result<SqliteAccess, rusqlite::Error> { 58 58 - let path = self.path.clone(); 112 112 + impl DiskStore { 113 113 + /// Initialize a new disk store 114 114 + pub async fn new( 115 115 + path: PathBuf, 116 116 + cache_mb: usize, 117 117 + max_stored_mb: usize, 118 118 + ) -> Result<Self, DiskError> { 119 119 + let max_stored = max_stored_mb * 2_usize.pow(20); 59 120 let conn = tokio::task::spawn_blocking(move || { 60 121 let conn = rusqlite::Connection::open(path)?; 61 122 123 123 + let sqlite_one_mb = -(2_i64.pow(10)); // negative is kibibytes for sqlite cache_size 124 124 + 125 125 + // conn.pragma_update(None, "journal_mode", "OFF")?; 126 126 + // conn.pragma_update(None, "journal_mode", "MEMORY")?; 62 127 conn.pragma_update(None, "journal_mode", "WAL")?; 128 128 + // conn.pragma_update(None, "wal_autocheckpoint", "0")?; // this lets things get a bit big on disk 63 129 conn.pragma_update(None, "synchronous", "OFF")?; 64 64 - conn.pragma_update(None, "cache_size", (-32 * 2_i64.pow(10)).to_string())?; 65 65 - conn.execute( 66 66 - "CREATE TABLE blocks ( 67 67 - key BLOB PRIMARY KEY NOT NULL, 68 68 - val BLOB NOT NULL 69 69 - ) WITHOUT ROWID", 70 70 - (), 130 130 + conn.pragma_update( 131 131 + None, 132 132 + "cache_size", 133 133 + (cache_mb as i64 * sqlite_one_mb).to_string(), 71 134 )?; 135 135 + Self::reset_tables(&conn)?; 72 136 73 73 - Ok::<_, Self::StorageError>(conn) 137 137 + Ok::<_, DiskError>(conn) 74 138 }) 75 75 - .await 76 76 - .expect("join error")?; 139 139 + .await??; 77 140 78 78 - Ok(SqliteAccess { conn }) 141 141 + Ok(Self { 142 142 + conn, 143 143 + max_stored, 144 144 + stored: 0, 145 145 + }) 79 146 } 80 80 - } 81 81 - 82 82 - pub struct SqliteAccess { 83 83 - conn: rusqlite::Connection, 84 84 - } 85 85 - 86 86 - impl DiskAccess for SqliteAccess { 87 87 - type StorageError = rusqlite::Error; 88 88 - fn get_writer(&mut self) -> Result<impl DiskWriter<rusqlite::Error>, rusqlite::Error> { 89 89 - let insert_stmt = self 90 90 - .conn 91 91 - .prepare("INSERT INTO blocks (key, val) VALUES (?1, ?2)")?; 92 92 - Ok(SqliteWriter { insert_stmt }) 147 147 + pub(crate) fn get_writer(&'_ mut self) -> Result<SqliteWriter<'_>, DiskError> { 148 148 + let tx = self.conn.transaction()?; 149 149 + Ok(SqliteWriter { 150 150 + tx, 151 151 + stored: &mut self.stored, 152 152 + max: self.max_stored, 153 153 + }) 93 154 } 94 94 - fn get_reader( 95 95 - &self, 96 96 - ) -> Result<impl DiskReader<StorageError = rusqlite::Error>, rusqlite::Error> { 155 155 + pub(crate) fn get_reader<'conn>(&'conn self) -> Result<SqliteReader<'conn>, DiskError> { 97 156 let select_stmt = self.conn.prepare("SELECT val FROM blocks WHERE key = ?1")?; 98 157 Ok(SqliteReader { select_stmt }) 99 158 } 159 159 + /// Drop and recreate the kv table 160 160 + pub async fn reset(self) -> Result<Self, DiskError> { 161 161 + tokio::task::spawn_blocking(move || { 162 162 + Self::reset_tables(&self.conn)?; 163 163 + Ok(self) 164 164 + }) 165 165 + .await? 166 166 + } 167 167 + fn reset_tables(conn: &rusqlite::Connection) -> Result<(), DiskError> { 168 168 + conn.execute("DROP TABLE IF EXISTS blocks", ())?; 169 169 + conn.execute( 170 170 + "CREATE TABLE blocks ( 171 171 + key BLOB PRIMARY KEY NOT NULL, 172 172 + val BLOB NOT NULL 173 173 + ) WITHOUT ROWID", 174 174 + (), 175 175 + )?; 176 176 + Ok(()) 177 177 + } 100 178 } 101 179 102 102 - pub struct SqliteWriter<'conn> { 103 103 - insert_stmt: rusqlite::Statement<'conn>, 180 180 + pub(crate) struct SqliteWriter<'conn> { 181 181 + tx: rusqlite::Transaction<'conn>, 182 182 + stored: &'conn mut usize, 183 183 + max: usize, 104 184 } 105 185 106 106 - impl DiskWriter<rusqlite::Error> for SqliteWriter<'_> { 107 107 - fn put(&mut self, key: Vec<u8>, val: Vec<u8>) -> rusqlite::Result<()> { 108 108 - self.insert_stmt.execute((key, val))?; 186 186 + impl SqliteWriter<'_> { 187 187 + pub(crate) fn put_many( 188 188 + &mut self, 189 189 + kv: impl Iterator<Item = Result<(Vec<u8>, Vec<u8>), DriveError>>, 190 190 + ) -> Result<(), DriveError> { 191 191 + let mut insert_stmt = self 192 192 + .tx 193 193 + .prepare_cached("INSERT INTO blocks (key, val) VALUES (?1, ?2)") 194 194 + .map_err(DiskError::DbError)?; 195 195 + for pair in kv { 196 196 + let (k, v) = pair?; 197 197 + *self.stored += v.len(); 198 198 + if *self.stored > self.max { 199 199 + return Err(DiskError::MaxSizeExceeded.into()); 200 200 + } 201 201 + insert_stmt.execute((k, v)).map_err(DiskError::DbError)?; 202 202 + } 203 203 + Ok(()) 204 204 + } 205 205 + pub fn commit(self) -> Result<(), DiskError> { 206 206 + self.tx.commit()?; 109 207 Ok(()) 110 208 } 111 209 } 112 210 113 113 - pub struct SqliteReader<'conn> { 211 211 + pub(crate) struct SqliteReader<'conn> { 114 212 select_stmt: rusqlite::Statement<'conn>, 115 213 } 116 214 117 117 - impl DiskReader for SqliteReader<'_> { 118 118 - type StorageError = rusqlite::Error; 119 119 - fn get(&mut self, key: Vec<u8>) -> rusqlite::Result<Option<Vec<u8>>> { 215 215 + impl SqliteReader<'_> { 216 216 + pub(crate) fn get(&mut self, key: Vec<u8>) -> rusqlite::Result<Option<Vec<u8>>> { 120 217 self.select_stmt 121 218 .query_one((&key,), |row| row.get(0)) 122 219 .optional() 123 220 } 124 221 } 125 125 - 126 126 - //////////// redb why not 127 127 - 128 128 - const REDB_TABLE: redb::TableDefinition<&[u8], &[u8]> = redb::TableDefinition::new("blocks"); 129 129 - 130 130 - pub struct RedbStore { 131 131 - path: PathBuf, 132 132 - } 133 133 - 134 134 - impl RedbStore { 135 135 - pub fn new(path: PathBuf) -> Self { 136 136 - Self { path } 137 137 - } 138 138 - } 139 139 - 140 140 - impl StorageErrorBase for redb::Error {} 141 141 - 142 142 - impl DiskStore for RedbStore { 143 143 - type StorageError = redb::Error; 144 144 - type Access = RedbAccess; 145 145 - async fn get_access(&mut self) -> Result<RedbAccess, redb::Error> { 146 146 - let path = self.path.clone(); 147 147 - let db = tokio::task::spawn_blocking(move || { 148 148 - let db = redb::Database::create(path)?; 149 149 - Ok::<_, Self::StorageError>(db) 150 150 - }) 151 151 - .await 152 152 - .expect("join error")?; 153 153 - 154 154 - Ok(RedbAccess { db }) 155 155 - } 156 156 - } 157 157 - 158 158 - pub struct RedbAccess { 159 159 - db: redb::Database, 160 160 - } 161 161 - 162 162 - impl DiskAccess for RedbAccess { 163 163 - type StorageError = redb::Error; 164 164 - fn get_writer(&mut self) -> Result<impl DiskWriter<redb::Error>, redb::Error> { 165 165 - let mut tx = self.db.begin_write()?; 166 166 - tx.set_durability(redb::Durability::None)?; 167 167 - Ok(RedbWriter { tx: Some(tx) }) 168 168 - } 169 169 - fn get_reader(&self) -> Result<impl DiskReader<StorageError = redb::Error>, redb::Error> { 170 170 - let tx = self.db.begin_read()?; 171 171 - Ok(RedbReader { tx }) 172 172 - } 173 173 - } 174 174 - 175 175 - pub struct RedbWriter { 176 176 - tx: Option<redb::WriteTransaction>, 177 177 - } 178 178 - 179 179 - impl DiskWriter<redb::Error> for RedbWriter { 180 180 - fn put(&mut self, key: Vec<u8>, val: Vec<u8>) -> Result<(), redb::Error> { 181 181 - let mut table = self.tx.as_ref().unwrap().open_table(REDB_TABLE)?; 182 182 - table.insert(&*key, &*val)?; 183 183 - Ok(()) 184 184 - } 185 185 - } 186 186 - 187 187 - /// oops careful in async 188 188 - impl Drop for RedbWriter { 189 189 - fn drop(&mut self) { 190 190 - let tx = self.tx.take(); 191 191 - tx.unwrap().commit().unwrap(); 192 192 - } 193 193 - } 194 194 - 195 195 - pub struct RedbReader { 196 196 - tx: redb::ReadTransaction, 197 197 - } 198 198 - 199 199 - impl DiskReader for RedbReader { 200 200 - type StorageError = redb::Error; 201 201 - fn get(&mut self, key: Vec<u8>) -> Result<Option<Vec<u8>>, redb::Error> { 202 202 - let table = self.tx.open_table(REDB_TABLE)?; 203 203 - let rv = table.get(&*key)?.map(|guard| guard.value().to_vec()); 204 204 - Ok(rv) 205 205 - } 206 206 - }

-245

src/disk_drive.rs

··· 1 1 - use futures::Stream; 2 2 - use futures::TryStreamExt; 3 3 - use std::collections::VecDeque; 4 4 - use std::error::Error; 5 5 - 6 6 - use crate::disk_walk::{RkeyError, Trip, Walker}; 7 7 - use crate::mst::{Commit, Node}; 8 8 - 9 9 - use ipld_core::cid::Cid; 10 10 - use serde::{Deserialize, Serialize, de::DeserializeOwned}; 11 11 - 12 12 - /// Errors that can happen while consuming and emitting blocks and records 13 13 - #[derive(Debug, thiserror::Error)] 14 14 - pub enum DriveError { 15 15 - #[error("Failed to initialize CarReader: {0}")] 16 16 - CarReader(#[from] iroh_car::Error), 17 17 - #[error("Car block stream error: {0}")] 18 18 - CarBlockError(Box<dyn Error>), 19 19 - #[error("Failed to decode commit block: {0}")] 20 20 - BadCommit(Box<dyn Error>), 21 21 - #[error("The Commit block reference by the root was not found")] 22 22 - MissingCommit, 23 23 - #[error("The MST block {0} could not be found")] 24 24 - MissingBlock(Cid), 25 25 - #[error("Failed to walk the mst tree: {0}")] 26 26 - Tripped(#[from] Trip), 27 27 - #[error("whatever: {0}")] 28 28 - WalkingProblem(#[from] WalkError), 29 29 - #[error("whatever: {0}")] 30 30 - Boooooo(String), 31 31 - #[error("Error while encoding: {0}")] 32 32 - EncodingError(#[from] bincode::error::EncodeError), 33 33 - #[error("Error while decoding: {0}")] 34 34 - DecodingError(#[from] bincode::error::DecodeError), 35 35 - } 36 36 - 37 37 - /// Limited subset of errors that can happen while walking 38 38 - #[derive(Debug, thiserror::Error)] 39 39 - pub enum WalkError { 40 40 - #[error("The MST block {0} could not be found")] 41 41 - MissingBlock(Cid), 42 42 - #[error("Failed to walk the mst tree: {0}")] 43 43 - Tripped(#[from] Trip), 44 44 - } 45 45 - 46 46 - #[derive(Debug, thiserror::Error)] 47 47 - pub enum BlockStoreError { 48 48 - #[error("Error from the storage backend: {0}")] 49 49 - StorageBackend(Box<dyn Error + Send>), 50 50 - 51 51 - #[error(transparent)] 52 52 - RkeyError(#[from] RkeyError), 53 53 - 54 54 - // this should probably not be up here 55 55 - #[error("Failed to join tokio task: {0}")] 56 56 - JoinError(tokio::task::JoinError), 57 57 - 58 58 - #[error("Could not find block: {0}")] 59 59 - MissingBlock(Cid), 60 60 - } 61 61 - 62 62 - #[derive(Serialize, Deserialize)] 63 63 - pub enum MaybeProcessedBlock<T: Serialize> { 64 64 - Raw(Vec<u8>), 65 65 - Processed(T), 66 66 - } 67 67 - 68 68 - pub type Records = Vec<(String, Vec<u8>)>; 69 69 - 70 70 - /// Storage backend for caching large-repo blocks 71 71 - /// 72 72 - /// Since 73 73 - pub trait BlockStore { 74 74 - fn put_batch( 75 75 - &self, 76 76 - blocks: Vec<(Cid, Vec<u8>)>, 77 77 - ) -> impl Future<Output = Result<(), BlockStoreError>>; // unwraps for now 78 78 - fn walk_batch( 79 79 - &self, 80 80 - walker: Walker, 81 81 - n: usize, 82 82 - ) -> impl Future<Output = Result<(Walker, Records), BlockStoreError>>; // boo string error for now because 83 83 - } 84 84 - 85 85 - type CarBlock<E> = Result<(Cid, Vec<u8>), E>; 86 86 - 87 87 - /// The core driver between the block stream and MST walker 88 88 - pub struct Vehicle<SE, S, T, BS, P> 89 89 - where 90 90 - SE: Error + 'static, 91 91 - S: Stream<Item = CarBlock<SE>>, 92 92 - T: Clone + Serialize + DeserializeOwned, 93 93 - BS: BlockStore, 94 94 - P: Fn(&[u8]) -> T, 95 95 - { 96 96 - #[allow(dead_code)] 97 97 - block_stream: Option<S>, 98 98 - block_store: BS, 99 99 - walker: Walker, 100 100 - process: P, 101 101 - out_cache: VecDeque<(String, T)>, 102 102 - } 103 103 - 104 104 - impl<SE, S, T, BS, P> Vehicle<SE, S, T, BS, P> 105 105 - where 106 106 - SE: Error + 'static, 107 107 - S: Stream<Item = CarBlock<SE>> + Unpin + Send, 108 108 - T: Clone + Serialize + DeserializeOwned + Send, 109 109 - BS: BlockStore + Send, 110 110 - P: Fn(&[u8]) -> T, 111 111 - { 112 112 - /// Set up the stream 113 113 - /// 114 114 - /// This will eagerly consume blocks until the `Commit` object is found. 115 115 - /// *Usually* the it's the first block, but there is no guarantee. 116 116 - /// 117 117 - /// ### Parameters 118 118 - /// 119 119 - /// `root`: CID of the commit object that is the root of the MST 120 120 - /// 121 121 - /// `block_stream`: Input stream of raw CAR blocks 122 122 - /// 123 123 - /// `process`: record-transforming callback: 124 124 - /// 125 125 - /// For tasks where records can be quickly processed into a *smaller* 126 126 - /// useful representation, you can do that eagerly as blocks come in by 127 127 - /// passing the processor as a callback here. This can reduce overall 128 128 - /// memory usage. 129 129 - pub async fn init( 130 130 - root: Cid, 131 131 - block_stream: S, 132 132 - block_store: BS, 133 133 - process: P, 134 134 - ) -> Result<(Commit, Self), DriveError> { 135 135 - let mut commit = None; 136 136 - 137 137 - log::warn!("init: load blocks"); 138 138 - 139 139 - let mut chunked = block_stream.try_chunks(256); 140 140 - 141 141 - // go ahead and put all blocks in the block store 142 142 - while let Some(chunk) = chunked 143 143 - .try_next() 144 144 - .await 145 145 - .map_err(|e| DriveError::CarBlockError(e.into()))? 146 146 - { 147 147 - let mut to_insert = Vec::with_capacity(chunk.len()); 148 148 - for (cid, data) in chunk { 149 149 - if cid == root { 150 150 - let c: Commit = serde_ipld_dagcbor::from_slice(&data) 151 151 - .map_err(|e| DriveError::BadCommit(e.into()))?; 152 152 - commit = Some(c); 153 153 - } else { 154 154 - let wrapped = if Node::could_be(&data) { 155 155 - MaybeProcessedBlock::Raw(data) 156 156 - } else { 157 157 - MaybeProcessedBlock::Processed(process(&data)) 158 158 - }; 159 159 - let bytes = 160 160 - bincode::serde::encode_to_vec(wrapped, bincode::config::standard())?; 161 161 - 162 162 - to_insert.push((cid, bytes)); 163 163 - } 164 164 - } 165 165 - block_store 166 166 - .put_batch(to_insert) 167 167 - .await 168 168 - .map_err(|e| DriveError::Boooooo(format!("boooOOOOO! {e}")))?; // TODO 169 169 - } 170 170 - 171 171 - log::warn!("init: got commit?"); 172 172 - 173 173 - // we either broke out or read all the blocks without finding the commit... 174 174 - let commit = commit.ok_or(DriveError::MissingCommit)?; 175 175 - 176 176 - let walker = Walker::new(commit.data); 177 177 - 178 178 - log::warn!("init: wrapping up"); 179 179 - 180 180 - let me = Self { 181 181 - block_stream: None, 182 182 - block_store, 183 183 - walker, 184 184 - process, 185 185 - out_cache: VecDeque::new(), 186 186 - }; 187 187 - Ok((commit, me)) 188 188 - } 189 189 - 190 190 - async fn load_chunk(&mut self, n: usize) -> Result<(), DriveError> { 191 191 - let walker = std::mem::take(&mut self.walker); 192 192 - let (walker, batch) = self 193 193 - .block_store 194 194 - .walk_batch(walker, n) 195 195 - .await 196 196 - .map_err(|e| DriveError::Boooooo(format!("booo! (here right?) {e}")))?; // TODO 197 197 - self.walker = walker; 198 198 - 199 199 - let processed = batch 200 200 - .into_iter() 201 201 - .map(|(k, encoded)| { 202 202 - let (decoded, n): (MaybeProcessedBlock<T>, usize) = 203 203 - bincode::serde::decode_from_slice(&encoded, bincode::config::standard())?; 204 204 - assert_eq!(n, encoded.len()); 205 205 - let processed = match decoded { 206 206 - MaybeProcessedBlock::Processed(t) => t, 207 207 - MaybeProcessedBlock::Raw(block) => (self.process)(&block), 208 208 - }; 209 209 - Ok((k, processed)) 210 210 - }) 211 211 - .collect::<Result<Vec<_>, DriveError>>()?; 212 212 - 213 213 - self.out_cache.extend(processed); 214 214 - Ok(()) 215 215 - } 216 216 - 217 217 - /// Get a chunk of records at a time 218 218 - /// 219 219 - /// the number of returned records may be smaller or larger than requested 220 220 - /// (but non-zero), even if it's not the last chunk. 221 221 - /// 222 222 - /// an empty vec will be returned to signal the end. 223 223 - pub async fn next_chunk(&mut self, n: usize) -> Result<Vec<(String, T)>, DriveError> { 224 224 - if self.out_cache.is_empty() { 225 225 - self.load_chunk(n).await?; 226 226 - } 227 227 - Ok(std::mem::take(&mut self.out_cache).into()) 228 228 - } 229 229 - 230 230 - /// Manually step through the record outputs 231 231 - pub async fn next_record(&mut self) -> Result<Option<(String, T)>, DriveError> { 232 232 - if self.out_cache.is_empty() { 233 233 - self.load_chunk(128).await?; // TODO 234 234 - } 235 235 - Ok(self.out_cache.pop_front()) 236 236 - } 237 237 - 238 238 - /// Convert to a futures::stream of record outputs 239 239 - pub fn stream(self) -> impl Stream<Item = Result<(String, T), DriveError>> { 240 240 - futures::stream::try_unfold(self, |mut this| async move { 241 241 - let maybe_record = this.next_record().await?; 242 242 - Ok(maybe_record.map(|b| (b, this))) 243 243 - }) 244 244 - } 245 245 - }

-110

src/disk_redb.rs

··· 1 1 - use crate::disk_drive::{BlockStore, BlockStoreError, MaybeProcessedBlock, Records}; 2 2 - use crate::disk_walk::{Need, Walker}; 3 3 - use ipld_core::cid::Cid; 4 4 - use redb::{Database, Durability, Error, ReadableDatabase, TableDefinition}; 5 5 - use std::path::Path; 6 6 - use std::sync::Arc; 7 7 - 8 8 - const TABLE: TableDefinition<&[u8], &[u8]> = TableDefinition::new("blocks"); 9 9 - 10 10 - pub struct RedbStore { 11 11 - #[allow(dead_code)] 12 12 - db: Arc<Database>, 13 13 - } 14 14 - 15 15 - impl RedbStore { 16 16 - pub async fn new(path: impl AsRef<Path> + 'static + Send) -> Result<Self, Error> { 17 17 - log::warn!("redb new"); 18 18 - let db = tokio::task::spawn_blocking(|| Database::create(path)) 19 19 - .await 20 20 - .unwrap()?; 21 21 - log::warn!("db created"); 22 22 - Ok(Self { db: db.into() }) 23 23 - } 24 24 - } 25 25 - 26 26 - // TODO: ship off to a blocking thread 27 27 - impl Drop for RedbStore { 28 28 - fn drop(&mut self) { 29 29 - let mut tx = self.db.begin_write().unwrap(); 30 30 - tx.set_durability(Durability::None).unwrap(); 31 31 - tx.delete_table(TABLE).unwrap(); 32 32 - tx.commit().unwrap(); 33 33 - } 34 34 - } 35 35 - 36 36 - impl<E: Into<Error>> From<E> for BlockStoreError { 37 37 - fn from(e: E) -> BlockStoreError { 38 38 - let e = Into::<Error>::into(e); 39 39 - BlockStoreError::StorageBackend(Box::new(e)) 40 40 - } 41 41 - } 42 42 - 43 43 - impl BlockStore for RedbStore { 44 44 - async fn put_batch(&self, blocks: Vec<(Cid, Vec<u8>)>) -> Result<(), BlockStoreError> { 45 45 - let db = self.db.clone(); 46 46 - tokio::task::spawn_blocking(move || -> Result<(), BlockStoreError> { 47 47 - let mut tx = db.begin_write()?; 48 48 - tx.set_durability(Durability::None)?; 49 49 - 50 50 - { 51 51 - let mut table = tx.open_table(TABLE)?; 52 52 - for (cid, t) in blocks { 53 53 - let key_bytes = cid.to_bytes(); 54 54 - table.insert(&*key_bytes, &*t)?; 55 55 - } 56 56 - } 57 57 - 58 58 - Ok(tx.commit()?) 59 59 - }) 60 60 - .await 61 61 - .map_err(BlockStoreError::JoinError)? 62 62 - } 63 63 - 64 64 - async fn walk_batch( 65 65 - &self, 66 66 - mut walker: Walker, 67 67 - n: usize, 68 68 - ) -> Result<(Walker, Records), BlockStoreError> { 69 69 - let db = self.db.clone(); 70 70 - tokio::task::spawn_blocking(move || -> Result<_, BlockStoreError> { 71 71 - let tx = db.begin_read()?; 72 72 - let table = tx.open_table(TABLE)?; 73 73 - 74 74 - let mut out = Vec::with_capacity(n); 75 75 - loop { 76 76 - let Some(need) = walker.next_needed()? else { 77 77 - break; 78 78 - }; 79 79 - let cid = need.cid(); 80 80 - let Some(res) = table.get(&*cid.to_bytes())? else { 81 81 - return Err(BlockStoreError::MissingBlock(cid)); 82 82 - }; 83 83 - let block = res.value(); 84 84 - 85 85 - match need { 86 86 - Need::Node(_) => { 87 87 - let (mpb, n) = 88 88 - bincode::serde::decode_from_slice(block, bincode::config::standard()) 89 89 - .unwrap(); 90 90 - assert_eq!(n, block.len()); 91 91 - // DANGER: we're throwing in unit () as a placeholder here and assuming bincode will still work since Raw is the first variant 92 92 - let MaybeProcessedBlock::Raw(bytes): MaybeProcessedBlock<()> = mpb else { 93 93 - panic!("should have not been processed"); // tODO 94 94 - }; 95 95 - walker.handle_node(&bytes)? 96 96 - } 97 97 - Need::Record { rkey, .. } => { 98 98 - out.push((rkey, block.to_vec())); 99 99 - if out.len() >= n { 100 100 - break; 101 101 - } 102 102 - } 103 103 - } 104 104 - } 105 105 - Ok((walker, out)) 106 106 - }) 107 107 - .await 108 108 - .map_err(BlockStoreError::JoinError)? 109 109 - } 110 110 - }

-65

src/disk_sqlite.rs

··· 1 1 - // use crate::disk_drive::BlockStore; 2 2 - // use ipld_core::cid::Cid; 3 3 - // use rusqlite::{Connection, OptionalExtension, Result}; 4 4 - // use serde::{Serialize, de::DeserializeOwned}; 5 5 - // use std::path::Path; 6 6 - 7 7 - // pub struct SqliteStore { 8 8 - // conn: Connection, 9 9 - // } 10 10 - 11 11 - // impl SqliteStore { 12 12 - // pub fn new(path: impl AsRef<Path>) -> Result<Self> { 13 13 - // let conn = Connection::open(path)?; 14 14 - // conn.pragma_update(None, "journal_mode", "WAL")?; 15 15 - // conn.pragma_update(None, "synchronous", "OFF")?; 16 16 - // conn.pragma_update(None, "cache_size", (-32 * 2_i64.pow(10)).to_string())?; 17 17 - // conn.execute( 18 18 - // "CREATE TABLE blocks ( 19 19 - // key BLOB PRIMARY KEY NOT NULL, 20 20 - // val BLOB NOT NULL 21 21 - // ) WITHOUT ROWID", 22 22 - // (), 23 23 - // )?; 24 24 - 25 25 - // Ok(Self { conn }) 26 26 - // } 27 27 - // } 28 28 - 29 29 - // impl Drop for SqliteStore { 30 30 - // fn drop(&mut self) { 31 31 - // self.conn.execute("DROP TABLE blocks", ()).unwrap(); 32 32 - // } 33 33 - // } 34 34 - 35 35 - // impl<MPB: Serialize + DeserializeOwned> BlockStore<MPB> for SqliteStore { 36 36 - // fn put(&self, c: Cid, t: MPB) { 37 37 - // let key_bytes = c.to_bytes(); 38 38 - // let val_bytes = bincode::serde::encode_to_vec(t, bincode::config::standard()).unwrap(); 39 39 - 40 40 - // self.conn 41 41 - // .execute( 42 42 - // "INSERT INTO blocks (key, val) VALUES (?1, ?2)", 43 43 - // (&key_bytes, &val_bytes), 44 44 - // ) 45 45 - // .unwrap(); 46 46 - // } 47 47 - // fn get(&self, c: Cid) -> Option<MPB> { 48 48 - // let key_bytes = c.to_bytes(); 49 49 - 50 50 - // let val_bytes: Vec<u8> = self 51 51 - // .conn 52 52 - // .query_one( 53 53 - // "SELECT val FROM blocks WHERE key = ?1", 54 54 - // (&key_bytes,), 55 55 - // |row| row.get(0), 56 56 - // ) 57 57 - // .optional() 58 58 - // .unwrap()?; 59 59 - 60 60 - // let (t, n): (MPB, usize) = 61 61 - // bincode::serde::decode_from_slice(&val_bytes, bincode::config::standard()).unwrap(); 62 62 - // assert_eq!(val_bytes.len(), n); 63 63 - // Some(t) 64 64 - // } 65 65 - // }

-369

src/disk_walk.rs

··· 1 1 - //! Depth-first MST traversal 2 2 - 3 3 - use crate::mst::Node; 4 4 - use std::convert::Infallible; 5 5 - 6 6 - use ipld_core::cid::Cid; 7 7 - use serde::{Serialize, de::DeserializeOwned}; 8 8 - 9 9 - /// Errors that can happen while walking 10 10 - #[derive(Debug, thiserror::Error)] 11 11 - pub enum Trip { 12 12 - #[error("empty mst nodes are not allowed")] 13 13 - NodeEmpty, 14 14 - #[error("Failed to decode commit block: {0}")] 15 15 - BadCommit(serde_ipld_dagcbor::DecodeError<Infallible>), 16 16 - #[error("Action node error: {0}")] 17 17 - RkeyError(#[from] RkeyError), 18 18 - #[error("Process failed: {0}")] 19 19 - ProcessFailed(String), 20 20 - } 21 21 - 22 22 - /// Errors from invalid Rkeys 23 23 - #[derive(Debug, thiserror::Error)] 24 24 - pub enum RkeyError { 25 25 - #[error("Failed to compute an rkey due to invalid prefix_len")] 26 26 - EntryPrefixOutOfbounds, 27 27 - #[error("RKey was not utf-8")] 28 28 - EntryRkeyNotUtf8(#[from] std::string::FromUtf8Error), 29 29 - #[error("Encountered an rkey out of order while walking the MST")] 30 30 - RkeyOutOfOrder, 31 31 - #[error("Failed to decode node block: {0}")] 32 32 - NodeDecodeError(#[from] serde_ipld_dagcbor::DecodeError<Infallible>), 33 33 - } 34 34 - 35 35 - /// Walker outputs 36 36 - #[derive(Debug)] 37 37 - pub enum Step<T: Serialize + DeserializeOwned> { 38 38 - /// We need a CID but it's not in the block store 39 39 - /// 40 40 - /// Give the needed CID to the driver so it can load blocks until it's found 41 41 - Rest(Cid), 42 42 - /// Reached the end of the MST! yay! 43 43 - Finish, 44 44 - /// A record was found! 45 45 - Step { rkey: String, data: T }, 46 46 - } 47 47 - 48 48 - #[derive(Debug, Clone, PartialEq)] 49 49 - pub enum Need { 50 50 - Node(Cid), 51 51 - Record { rkey: String, cid: Cid }, 52 52 - } 53 53 - 54 54 - impl Need { 55 55 - pub fn cid(&self) -> Cid { 56 56 - match self { 57 57 - Need::Node(cid) => *cid, 58 58 - Need::Record { cid, .. } => *cid, 59 59 - } 60 60 - } 61 61 - } 62 62 - 63 63 - fn push_from_node(stack: &mut Vec<Need>, node: &Node) -> Result<(), RkeyError> { 64 64 - let mut entries = Vec::with_capacity(node.entries.len()); 65 65 - 66 66 - let mut prefix = vec![]; 67 67 - for entry in &node.entries { 68 68 - let mut rkey = vec![]; 69 69 - let pre_checked = prefix 70 70 - .get(..entry.prefix_len) 71 71 - .ok_or(RkeyError::EntryPrefixOutOfbounds)?; 72 72 - rkey.extend_from_slice(pre_checked); 73 73 - rkey.extend_from_slice(&entry.keysuffix); 74 74 - prefix = rkey.clone(); 75 75 - 76 76 - entries.push(Need::Record { 77 77 - rkey: String::from_utf8(rkey)?, 78 78 - cid: entry.value, 79 79 - }); 80 80 - if let Some(ref tree) = entry.tree { 81 81 - entries.push(Need::Node(*tree)); 82 82 - } 83 83 - } 84 84 - 85 85 - entries.reverse(); 86 86 - stack.append(&mut entries); 87 87 - 88 88 - if let Some(tree) = node.left { 89 89 - stack.push(Need::Node(tree)); 90 90 - } 91 91 - Ok(()) 92 92 - } 93 93 - 94 94 - /// Traverser of an atproto MST 95 95 - /// 96 96 - /// Walks the tree from left-to-right in depth-first order 97 97 - /// 98 98 - /// (turning into more of a navigator) 99 99 - /// it doesn't quite feel like the divisions of responsibility are right around 100 100 - /// here yet. 101 101 - #[derive(Debug, Default)] 102 102 - pub struct Walker { 103 103 - stack: Vec<Need>, 104 104 - prev: String, 105 105 - } 106 106 - 107 107 - impl Walker { 108 108 - pub fn new(tree_root_cid: Cid) -> Self { 109 109 - Self { 110 110 - stack: vec![Need::Node(tree_root_cid)], 111 111 - prev: "".to_string(), 112 112 - } 113 113 - } 114 114 - 115 115 - pub fn next_needed(&mut self) -> Result<Option<Need>, RkeyError> { 116 116 - let Some(need) = self.stack.pop() else { 117 117 - return Ok(None); 118 118 - }; 119 119 - if let Need::Record { ref rkey, .. } = need { 120 120 - // rkeys *must* be in order or else the tree is invalid (or 121 121 - // we have a bug) 122 122 - if *rkey <= self.prev { 123 123 - return Err(RkeyError::RkeyOutOfOrder); 124 124 - } 125 125 - self.prev = rkey.clone(); 126 126 - } 127 127 - Ok(Some(need)) 128 128 - } 129 129 - 130 130 - /// hacky: this must be called after next_needed if it was a node 131 131 - pub fn handle_node(&mut self, block: &[u8]) -> Result<(), RkeyError> { 132 132 - let node = serde_ipld_dagcbor::from_slice::<Node>(block)?; 133 133 - push_from_node(&mut self.stack, &node)?; 134 134 - Ok(()) 135 135 - } 136 136 - } 137 137 - 138 138 - #[cfg(test)] 139 139 - mod test { 140 140 - use super::*; 141 141 - // use crate::mst::Entry; 142 142 - 143 143 - fn cid1() -> Cid { 144 144 - "bafyreihixenvk3ahqbytas4hk4a26w43bh6eo3w6usjqtxkpzsvi655a3m" 145 145 - .parse() 146 146 - .unwrap() 147 147 - } 148 148 - // fn cid2() -> Cid { 149 149 - // "QmY7Yh4UquoXHLPFo2XbhXkhBvFoPwmQUSa92pxnxjQuPU" 150 150 - // .parse() 151 151 - // .unwrap() 152 152 - // } 153 153 - // fn cid3() -> Cid { 154 154 - // "bafybeigdyrzt5sfp7udm7hu76uh7y26nf3efuylqabf3oclgtqy55fbzdi" 155 155 - // .parse() 156 156 - // .unwrap() 157 157 - // } 158 158 - // fn cid4() -> Cid { 159 159 - // "QmbWqxBEKC3P8tqsKc98xmWNzrzDtRLMiMPL8wBuTGsMnR" 160 160 - // .parse() 161 161 - // .unwrap() 162 162 - // } 163 163 - // fn cid5() -> Cid { 164 164 - // "QmSnuWmxptJZdLJpKRarxBMS2Ju2oANVrgbr2xWbie9b2D" 165 165 - // .parse() 166 166 - // .unwrap() 167 167 - // } 168 168 - // fn cid6() -> Cid { 169 169 - // "QmdmQXB2mzChmMeKY47C43LxUdg1NDJ5MWcKMKxDu7RgQm" 170 170 - // .parse() 171 171 - // .unwrap() 172 172 - // } 173 173 - // fn cid7() -> Cid { 174 174 - // "bafybeiaysi4s6lnjev27ln5icwm6tueaw2vdykrtjkwiphwekaywqhcjze" 175 175 - // .parse() 176 176 - // .unwrap() 177 177 - // } 178 178 - // fn cid8() -> Cid { 179 179 - // "bafyreif3tfdpr5n4jdrbielmcapwvbpcthepfkwq2vwonmlhirbjmotedi" 180 180 - // .parse() 181 181 - // .unwrap() 182 182 - // } 183 183 - // fn cid9() -> Cid { 184 184 - // "bafyreicnokmhmrnlp2wjhyk2haep4tqxiptwfrp2rrs7rzq7uk766chqvq" 185 185 - // .parse() 186 186 - // .unwrap() 187 187 - // } 188 188 - 189 189 - #[test] 190 190 - fn test_next_from_node_empty() { 191 191 - let node = Node { 192 192 - left: None, 193 193 - entries: vec![], 194 194 - }; 195 195 - let mut stack = vec![]; 196 196 - push_from_node(&mut stack, &node).unwrap(); 197 197 - assert_eq!(stack.last(), None); 198 198 - } 199 199 - 200 200 - #[test] 201 201 - fn test_needs_from_node_just_left() { 202 202 - let node = Node { 203 203 - left: Some(cid1()), 204 204 - entries: vec![], 205 205 - }; 206 206 - let mut stack = vec![]; 207 207 - push_from_node(&mut stack, &node).unwrap(); 208 208 - assert_eq!(stack.last(), Some(Need::Node(cid1())).as_ref()); 209 209 - } 210 210 - 211 211 - // #[test] 212 212 - // fn test_needs_from_node_just_one_record() { 213 213 - // let node = Node { 214 214 - // left: None, 215 215 - // entries: vec![Entry { 216 216 - // keysuffix: "asdf".into(), 217 217 - // prefix_len: 0, 218 218 - // value: cid1(), 219 219 - // tree: None, 220 220 - // }], 221 221 - // }; 222 222 - // assert_eq!( 223 223 - // needs_from_node(node).unwrap(), 224 224 - // vec![Need::Record { 225 225 - // rkey: "asdf".into(), 226 226 - // cid: cid1(), 227 227 - // },] 228 228 - // ); 229 229 - // } 230 230 - 231 231 - // #[test] 232 232 - // fn test_needs_from_node_two_records() { 233 233 - // let node = Node { 234 234 - // left: None, 235 235 - // entries: vec![ 236 236 - // Entry { 237 237 - // keysuffix: "asdf".into(), 238 238 - // prefix_len: 0, 239 239 - // value: cid1(), 240 240 - // tree: None, 241 241 - // }, 242 242 - // Entry { 243 243 - // keysuffix: "gh".into(), 244 244 - // prefix_len: 2, 245 245 - // value: cid2(), 246 246 - // tree: None, 247 247 - // }, 248 248 - // ], 249 249 - // }; 250 250 - // assert_eq!( 251 251 - // needs_from_node(node).unwrap(), 252 252 - // vec![ 253 253 - // Need::Record { 254 254 - // rkey: "asdf".into(), 255 255 - // cid: cid1(), 256 256 - // }, 257 257 - // Need::Record { 258 258 - // rkey: "asgh".into(), 259 259 - // cid: cid2(), 260 260 - // }, 261 261 - // ] 262 262 - // ); 263 263 - // } 264 264 - 265 265 - // #[test] 266 266 - // fn test_needs_from_node_with_both() { 267 267 - // let node = Node { 268 268 - // left: None, 269 269 - // entries: vec![Entry { 270 270 - // keysuffix: "asdf".into(), 271 271 - // prefix_len: 0, 272 272 - // value: cid1(), 273 273 - // tree: Some(cid2()), 274 274 - // }], 275 275 - // }; 276 276 - // assert_eq!( 277 277 - // needs_from_node(node).unwrap(), 278 278 - // vec![ 279 279 - // Need::Record { 280 280 - // rkey: "asdf".into(), 281 281 - // cid: cid1(), 282 282 - // }, 283 283 - // Need::Node(cid2()), 284 284 - // ] 285 285 - // ); 286 286 - // } 287 287 - 288 288 - // #[test] 289 289 - // fn test_needs_from_node_left_and_record() { 290 290 - // let node = Node { 291 291 - // left: Some(cid1()), 292 292 - // entries: vec![Entry { 293 293 - // keysuffix: "asdf".into(), 294 294 - // prefix_len: 0, 295 295 - // value: cid2(), 296 296 - // tree: None, 297 297 - // }], 298 298 - // }; 299 299 - // assert_eq!( 300 300 - // needs_from_node(node).unwrap(), 301 301 - // vec![ 302 302 - // Need::Node(cid1()), 303 303 - // Need::Record { 304 304 - // rkey: "asdf".into(), 305 305 - // cid: cid2(), 306 306 - // }, 307 307 - // ] 308 308 - // ); 309 309 - // } 310 310 - 311 311 - // #[test] 312 312 - // fn test_needs_from_full_node() { 313 313 - // let node = Node { 314 314 - // left: Some(cid1()), 315 315 - // entries: vec![ 316 316 - // Entry { 317 317 - // keysuffix: "asdf".into(), 318 318 - // prefix_len: 0, 319 319 - // value: cid2(), 320 320 - // tree: Some(cid3()), 321 321 - // }, 322 322 - // Entry { 323 323 - // keysuffix: "ghi".into(), 324 324 - // prefix_len: 1, 325 325 - // value: cid4(), 326 326 - // tree: Some(cid5()), 327 327 - // }, 328 328 - // Entry { 329 329 - // keysuffix: "jkl".into(), 330 330 - // prefix_len: 2, 331 331 - // value: cid6(), 332 332 - // tree: Some(cid7()), 333 333 - // }, 334 334 - // Entry { 335 335 - // keysuffix: "mno".into(), 336 336 - // prefix_len: 4, 337 337 - // value: cid8(), 338 338 - // tree: Some(cid9()), 339 339 - // }, 340 340 - // ], 341 341 - // }; 342 342 - // assert_eq!( 343 343 - // needs_from_node(node).unwrap(), 344 344 - // vec![ 345 345 - // Need::Node(cid1()), 346 346 - // Need::Record { 347 347 - // rkey: "asdf".into(), 348 348 - // cid: cid2(), 349 349 - // }, 350 350 - // Need::Node(cid3()), 351 351 - // Need::Record { 352 352 - // rkey: "aghi".into(), 353 353 - // cid: cid4(), 354 354 - // }, 355 355 - // Need::Node(cid5()), 356 356 - // Need::Record { 357 357 - // rkey: "agjkl".into(), 358 358 - // cid: cid6(), 359 359 - // }, 360 360 - // Need::Node(cid7()), 361 361 - // Need::Record { 362 362 - // rkey: "agjkmno".into(), 363 363 - // cid: cid8(), 364 364 - // }, 365 365 - // Need::Node(cid9()), 366 366 - // ] 367 367 - // ); 368 368 - // } 369 369 - }

+485 -210

src/drive.rs

··· 1 1 - //! Consume an MST block stream, producing an ordered stream of records 1 1 + //! Consume a CAR from an AsyncRead, producing an ordered stream of records 2 2 3 3 - use crate::disk::{DiskAccess, DiskStore, DiskWriter, StorageErrorBase}; 3 3 + use crate::disk::{DiskError, DiskStore}; 4 4 + use crate::process::Processable; 4 5 use ipld_core::cid::Cid; 5 6 use iroh_car::CarReader; 6 6 - use serde::de::DeserializeOwned; 7 7 use serde::{Deserialize, Serialize}; 8 8 use std::collections::HashMap; 9 9 use std::convert::Infallible; 10 10 - use tokio::io::AsyncRead; 10 10 + use tokio::{io::AsyncRead, sync::mpsc}; 11 11 12 12 use crate::mst::{Commit, Node}; 13 13 - use crate::walk::{DiskTrip, Step, Trip, Walker}; 13 13 + use crate::walk::{Step, WalkError, Walker}; 14 14 15 15 /// Errors that can happen while consuming and emitting blocks and records 16 16 #[derive(Debug, thiserror::Error)] ··· 24 24 #[error("The MST block {0} could not be found")] 25 25 MissingBlock(Cid), 26 26 #[error("Failed to walk the mst tree: {0}")] 27 27 - Tripped(#[from] Trip), 27 27 + WalkError(#[from] WalkError), 28 28 #[error("CAR file had no roots")] 29 29 MissingRoot, 30 30 - } 31 31 - 32 32 - #[derive(Debug, thiserror::Error)] 33 33 - pub enum DiskDriveError<E: StorageErrorBase> { 34 34 - #[error("Error from iroh_car: {0}")] 35 35 - CarReader(#[from] iroh_car::Error), 36 36 - #[error("Failed to decode commit block: {0}")] 37 37 - BadBlock(#[from] serde_ipld_dagcbor::DecodeError<Infallible>), 38 30 #[error("Storage error")] 39 39 - StorageError(#[from] E), 40 40 - #[error("The Commit block reference by the root was not found")] 41 41 - MissingCommit, 42 42 - #[error("The MST block {0} could not be found")] 43 43 - MissingBlock(Cid), 31 31 + StorageError(#[from] DiskError), 44 32 #[error("Encode error: {0}")] 45 33 BincodeEncodeError(#[from] bincode::error::EncodeError), 46 46 - #[error("Decode error: {0}")] 34 34 + #[error("Tried to send on a closed channel")] 35 35 + ChannelSendError, // SendError takes <T> which we don't need 36 36 + #[error("Failed to join a task: {0}")] 37 37 + JoinError(#[from] tokio::task::JoinError), 38 38 + } 39 39 + 40 40 + #[derive(Debug, thiserror::Error)] 41 41 + pub enum DecodeError { 42 42 + #[error(transparent)] 47 43 BincodeDecodeError(#[from] bincode::error::DecodeError), 48 48 - #[error("disk tripped: {0}")] 49 49 - DiskTripped(#[from] DiskTrip<E>), 44 44 + #[error("extra bytes remained after decoding")] 45 45 + ExtraGarbage, 50 46 } 51 47 52 52 - // #[derive(Debug, thiserror::Error)] 53 53 - // pub enum Boooooo<E: StorageErrorBase> { 54 54 - // #[error("disk tripped: {0}")] 55 55 - // DiskTripped(#[from] DiskTrip<E>), 56 56 - // #[error("dde whatever: {0}")] 57 57 - // DiskDriveError(#[from] DiskDriveError<E>), 58 58 - // } 59 59 - 60 60 - pub trait Processable: Clone + Serialize + DeserializeOwned {} 48 48 + /// An in-order chunk of Rkey + (processed) Block pairs 49 49 + pub type BlockChunk<T> = Vec<(String, T)>; 61 50 62 51 #[derive(Debug, Clone, Serialize, Deserialize)] 63 63 - pub enum MaybeProcessedBlock<T> { 52 52 + pub(crate) enum MaybeProcessedBlock<T> { 64 53 /// A block that's *probably* a Node (but we can't know yet) 65 54 /// 66 55 /// It *can be* a record that suspiciously looks a lot like a node, so we ··· 85 74 Processed(T), 86 75 } 87 76 88 88 - impl<T: Processable> Processable for MaybeProcessedBlock<T> {} 77 77 + impl<T: Processable> Processable for MaybeProcessedBlock<T> { 78 78 + /// TODO this is probably a little broken 79 79 + fn get_size(&self) -> usize { 80 80 + use std::{cmp::max, mem::size_of}; 81 81 + 82 82 + // enum is always as big as its biggest member? 83 83 + let base_size = max(size_of::<Vec<u8>>(), size_of::<T>()); 84 84 + 85 85 + let extra = match self { 86 86 + Self::Raw(bytes) => bytes.len(), 87 87 + Self::Processed(t) => t.get_size(), 88 88 + }; 89 89 90 90 - pub enum Vehicle<R: AsyncRead + Unpin, T: Processable> { 91 91 - Lil(Commit, MemDriver<T>), 92 92 - Big(BigCar<R, T>), 90 90 + base_size + extra 91 91 + } 93 92 } 94 93 95 95 - pub async fn load_car<R: AsyncRead + Unpin, T: Processable>( 96 96 - reader: R, 97 97 - process: fn(&[u8]) -> T, 98 98 - max_size: usize, 99 99 - ) -> Result<Vehicle<R, T>, DriveError> { 100 100 - let mut mem_blocks = HashMap::new(); 94 94 + impl<T> MaybeProcessedBlock<T> { 95 95 + fn maybe(process: fn(Vec<u8>) -> T, data: Vec<u8>) -> Self { 96 96 + if Node::could_be(&data) { 97 97 + MaybeProcessedBlock::Raw(data) 98 98 + } else { 99 99 + MaybeProcessedBlock::Processed(process(data)) 100 100 + } 101 101 + } 102 102 + } 101 103 102 102 - let mut car = CarReader::new(reader).await?; 104 104 + /// Read a CAR file, buffering blocks in memory or to disk 105 105 + pub enum Driver<R: AsyncRead + Unpin, T: Processable> { 106 106 + /// All blocks fit within the memory limit 107 107 + /// 108 108 + /// You probably want to check the commit's signature. You can go ahead and 109 109 + /// walk the MST right away. 110 110 + Memory(Commit, MemDriver<T>), 111 111 + /// Blocks exceed the memory limit 112 112 + /// 113 113 + /// You'll need to provide a disk storage to continue. The commit will be 114 114 + /// returned and can be validated only once all blocks are loaded. 115 115 + Disk(NeedDisk<R, T>), 116 116 + } 103 117 104 104 - let root = *car 105 105 - .header() 106 106 - .roots() 107 107 - .first() 108 108 - .ok_or(DriveError::MissingRoot)?; 109 109 - log::debug!("root: {root:?}"); 118 118 + /// Builder-style driver setup 119 119 + #[derive(Debug, Clone)] 120 120 + pub struct DriverBuilder { 121 121 + pub mem_limit_mb: usize, 122 122 + } 110 123 111 111 - let mut commit = None; 124 124 + impl Default for DriverBuilder { 125 125 + fn default() -> Self { 126 126 + Self { mem_limit_mb: 16 } 127 127 + } 128 128 + } 112 129 113 113 - // try to load all the blocks into memory 114 114 - while let Some((cid, data)) = car.next_block().await? { 115 115 - // the root commit is a Special Third Kind of block that we need to make 116 116 - // sure not to optimistically send to the processing function 117 117 - if cid == root { 118 118 - let c: Commit = serde_ipld_dagcbor::from_slice(&data)?; 119 119 - commit = Some(c); 120 120 - continue; 130 130 + impl DriverBuilder { 131 131 + /// Begin configuring the driver with defaults 132 132 + pub fn new() -> Self { 133 133 + Default::default() 134 134 + } 135 135 + /// Set the in-memory size limit, in MiB 136 136 + /// 137 137 + /// Default: 16 MiB 138 138 + pub fn with_mem_limit_mb(self, new_limit: usize) -> Self { 139 139 + Self { 140 140 + mem_limit_mb: new_limit, 141 141 + } 142 142 + } 143 143 + /// Set the block processor 144 144 + /// 145 145 + /// Default: noop, raw blocks will be emitted 146 146 + pub fn with_block_processor<T: Processable>( 147 147 + self, 148 148 + p: fn(Vec<u8>) -> T, 149 149 + ) -> DriverBuilderWithProcessor<T> { 150 150 + DriverBuilderWithProcessor { 151 151 + mem_limit_mb: self.mem_limit_mb, 152 152 + block_processor: p, 121 153 } 154 154 + } 155 155 + /// Begin processing an atproto MST from a CAR file 156 156 + pub async fn load_car<R: AsyncRead + Unpin>( 157 157 + &self, 158 158 + reader: R, 159 159 + ) -> Result<Driver<R, Vec<u8>>, DriveError> { 160 160 + Driver::load_car(reader, crate::process::noop, self.mem_limit_mb).await 161 161 + } 162 162 + } 122 163 123 123 - // remaining possible types: node, record, other. optimistically process 124 124 - // TODO: get the actual in-memory size to compute disk spill 125 125 - let maybe_processed = if Node::could_be(&data) { 126 126 - MaybeProcessedBlock::Raw(data) 127 127 - } else { 128 128 - MaybeProcessedBlock::Processed(process(&data)) 129 129 - }; 164 164 + /// Builder-style driver intermediate step 165 165 + /// 166 166 + /// start from `DriverBuilder` 167 167 + #[derive(Debug, Clone)] 168 168 + pub struct DriverBuilderWithProcessor<T: Processable> { 169 169 + pub mem_limit_mb: usize, 170 170 + pub block_processor: fn(Vec<u8>) -> T, 171 171 + } 130 172 131 131 - // stash (maybe processed) blocks in memory as long as we have room 132 132 - mem_blocks.insert(cid, maybe_processed); 133 133 - if mem_blocks.len() >= max_size { 134 134 - return Ok(Vehicle::Big(BigCar { 135 135 - car, 136 136 - root, 137 137 - process, 138 138 - max_size, 139 139 - mem_blocks, 140 140 - commit, 141 141 - })); 173 173 + impl<T: Processable> DriverBuilderWithProcessor<T> { 174 174 + /// Set the in-memory size limit, in MiB 175 175 + /// 176 176 + /// Default: 16 MiB 177 177 + pub fn with_mem_limit_mb(mut self, new_limit: usize) -> Self { 178 178 + self.mem_limit_mb = new_limit; 179 179 + self 180 180 + } 181 181 + /// Begin processing an atproto MST from a CAR file 182 182 + pub async fn load_car<R: AsyncRead + Unpin>( 183 183 + &self, 184 184 + reader: R, 185 185 + ) -> Result<Driver<R, T>, DriveError> { 186 186 + Driver::load_car(reader, self.block_processor, self.mem_limit_mb).await 187 187 + } 188 188 + } 189 189 + 190 190 + impl<R: AsyncRead + Unpin, T: Processable> Driver<R, T> { 191 191 + /// Begin processing an atproto MST from a CAR file 192 192 + /// 193 193 + /// Blocks will be loaded, processed, and buffered in memory. If the entire 194 194 + /// processed size is under the `mem_limit_mb` limit, a `Driver::Memory` 195 195 + /// will be returned along with a `Commit` ready for validation. 196 196 + /// 197 197 + /// If the `mem_limit_mb` limit is reached before loading all blocks, the 198 198 + /// partial state will be returned as `Driver::Disk(needed)`, which can be 199 199 + /// resumed by providing a `SqliteStorage` for on-disk block storage. 200 200 + pub async fn load_car( 201 201 + reader: R, 202 202 + process: fn(Vec<u8>) -> T, 203 203 + mem_limit_mb: usize, 204 204 + ) -> Result<Driver<R, T>, DriveError> { 205 205 + let max_size = mem_limit_mb * 2_usize.pow(20); 206 206 + let mut mem_blocks = HashMap::new(); 207 207 + 208 208 + let mut car = CarReader::new(reader).await?; 209 209 + 210 210 + let root = *car 211 211 + .header() 212 212 + .roots() 213 213 + .first() 214 214 + .ok_or(DriveError::MissingRoot)?; 215 215 + log::debug!("root: {root:?}"); 216 216 + 217 217 + let mut commit = None; 218 218 + 219 219 + // try to load all the blocks into memory 220 220 + let mut mem_size = 0; 221 221 + while let Some((cid, data)) = car.next_block().await? { 222 222 + // the root commit is a Special Third Kind of block that we need to make 223 223 + // sure not to optimistically send to the processing function 224 224 + if cid == root { 225 225 + let c: Commit = serde_ipld_dagcbor::from_slice(&data)?; 226 226 + commit = Some(c); 227 227 + continue; 228 228 + } 229 229 + 230 230 + // remaining possible types: node, record, other. optimistically process 231 231 + let maybe_processed = MaybeProcessedBlock::maybe(process, data); 232 232 + 233 233 + // stash (maybe processed) blocks in memory as long as we have room 234 234 + mem_size += std::mem::size_of::<Cid>() + maybe_processed.get_size(); 235 235 + mem_blocks.insert(cid, maybe_processed); 236 236 + if mem_size >= max_size { 237 237 + return Ok(Driver::Disk(NeedDisk { 238 238 + car, 239 239 + root, 240 240 + process, 241 241 + max_size, 242 242 + mem_blocks, 243 243 + commit, 244 244 + })); 245 245 + } 142 246 } 247 247 + 248 248 + // all blocks loaded and we fit in memory! hopefully we found the commit... 249 249 + let commit = commit.ok_or(DriveError::MissingCommit)?; 250 250 + 251 251 + let walker = Walker::new(commit.data); 252 252 + 253 253 + Ok(Driver::Memory( 254 254 + commit, 255 255 + MemDriver { 256 256 + blocks: mem_blocks, 257 257 + walker, 258 258 + process, 259 259 + }, 260 260 + )) 143 261 } 262 262 + } 144 263 145 145 - // all blocks loaded and we fit in memory! hopefully we found the commit... 146 146 - let commit = commit.ok_or(DriveError::MissingCommit)?; 264 264 + /// The core driver between the block stream and MST walker 265 265 + /// 266 266 + /// In the future, PDSs will export CARs in a stream-friendly order that will 267 267 + /// enable processing them with tiny memory overhead. But that future is not 268 268 + /// here yet. 269 269 + /// 270 270 + /// CARs are almost always in a stream-unfriendly order, so I'm reverting the 271 271 + /// optimistic stream features: we load all block first, then walk the MST. 272 272 + /// 273 273 + /// This makes things much simpler: we only need to worry about spilling to disk 274 274 + /// in one place, and we always have a reasonable expecatation about how much 275 275 + /// work the init function will do. We can drop the CAR reader before walking, 276 276 + /// so the sync/async boundaries become a little easier to work around. 277 277 + #[derive(Debug)] 278 278 + pub struct MemDriver<T: Processable> { 279 279 + blocks: HashMap<Cid, MaybeProcessedBlock<T>>, 280 280 + walker: Walker, 281 281 + process: fn(Vec<u8>) -> T, 282 282 + } 147 283 148 148 - let walker = Walker::new(commit.data); 284 284 + impl<T: Processable> MemDriver<T> { 285 285 + /// Step through the record outputs, in rkey order 286 286 + pub async fn next_chunk(&mut self, n: usize) -> Result<Option<BlockChunk<T>>, DriveError> { 287 287 + let mut out = Vec::with_capacity(n); 288 288 + for _ in 0..n { 289 289 + // walk as far as we can until we run out of blocks or find a record 290 290 + match self.walker.step(&mut self.blocks, self.process)? { 291 291 + Step::Missing(cid) => return Err(DriveError::MissingBlock(cid)), 292 292 + Step::Finish => break, 293 293 + Step::Found { rkey, data } => { 294 294 + out.push((rkey, data)); 295 295 + continue; 296 296 + } 297 297 + }; 298 298 + } 149 299 150 150 - Ok(Vehicle::Lil( 151 151 - commit, 152 152 - MemDriver { 153 153 - blocks: mem_blocks, 154 154 - walker, 155 155 - process, 156 156 - }, 157 157 - )) 300 300 + if out.is_empty() { 301 301 + Ok(None) 302 302 + } else { 303 303 + Ok(Some(out)) 304 304 + } 305 305 + } 158 306 } 159 307 160 160 - /// a paritally memory-loaded car file that needs disk spillover to continue 161 161 - pub struct BigCar<R: AsyncRead + Unpin, T: Processable> { 308 308 + /// A partially memory-loaded car file that needs disk spillover to continue 309 309 + pub struct NeedDisk<R: AsyncRead + Unpin, T: Processable> { 162 310 car: CarReader<R>, 163 311 root: Cid, 164 164 - process: fn(&[u8]) -> T, 312 312 + process: fn(Vec<u8>) -> T, 165 313 max_size: usize, 166 314 mem_blocks: HashMap<Cid, MaybeProcessedBlock<T>>, 167 315 pub commit: Option<Commit>, ··· 171 319 bincode::serde::encode_to_vec(v, bincode::config::standard()) 172 320 } 173 321 174 174 - pub fn decode<T: Processable>(bytes: &[u8]) -> Result<T, bincode::error::DecodeError> { 322 322 + pub(crate) fn decode<T: Processable>(bytes: &[u8]) -> Result<T, DecodeError> { 175 323 let (t, n) = bincode::serde::decode_from_slice(bytes, bincode::config::standard())?; 176 176 - assert_eq!(n, bytes.len(), "expected to decode all bytes"); // TODO 324 324 + if n != bytes.len() { 325 325 + return Err(DecodeError::ExtraGarbage); 326 326 + } 177 327 Ok(t) 178 328 } 179 329 180 180 - impl<R: AsyncRead + Unpin, T: Processable + Send + 'static> BigCar<R, T> { 181 181 - pub async fn finish_loading<S: DiskStore>( 330 330 + impl<R: AsyncRead + Unpin, T: Processable + Send + 'static> NeedDisk<R, T> { 331 331 + pub async fn finish_loading( 182 332 mut self, 183 183 - mut store: S, 184 184 - ) -> Result<(Commit, BigCarReady<T, S::Access>), DiskDriveError<S::StorageError>> 185 185 - where 186 186 - S::Access: Send + 'static, 187 187 - S::StorageError: 'static, 188 188 - { 189 189 - // set up access for real 190 190 - let mut access = store.get_access().await?; 333 333 + mut store: DiskStore, 334 334 + ) -> Result<(Commit, DiskDriver<T>), DriveError> { 335 335 + // move store in and back out so we can manage lifetimes 336 336 + // dump mem blocks into the store 337 337 + store = tokio::task::spawn(async move { 338 338 + let mut writer = store.get_writer()?; 339 339 + 340 340 + let kvs = self 341 341 + .mem_blocks 342 342 + .into_iter() 343 343 + .map(|(k, v)| Ok(encode(v).map(|v| (k.to_bytes(), v))?)); 191 344 192 192 - // move access in and back out so we can manage lifetimes 193 193 - // dump mem blocks into the store 194 194 - access = tokio::task::spawn(async move { 195 195 - let mut writer = access.get_writer()?; 196 196 - for (k, v) in self.mem_blocks { 197 197 - let key_bytes = k.to_bytes(); 198 198 - let val_bytes = encode(v)?; // TODO 199 199 - writer.put(key_bytes, val_bytes)?; 200 200 - } 201 201 - drop(writer); // cannot outlive access 202 202 - Ok::<_, DiskDriveError<S::StorageError>>(access) 345 345 + writer.put_many(kvs)?; 346 346 + writer.commit()?; 347 347 + Ok::<_, DriveError>(store) 203 348 }) 204 204 - .await 205 205 - .unwrap()?; 349 349 + .await??; 350 350 + 351 351 + let (tx, mut rx) = mpsc::channel::<Vec<(Cid, MaybeProcessedBlock<T>)>>(1); 352 352 + 353 353 + let store_worker = tokio::task::spawn_blocking(move || { 354 354 + let mut writer = store.get_writer()?; 355 355 + 356 356 + while let Some(chunk) = rx.blocking_recv() { 357 357 + let kvs = chunk 358 358 + .into_iter() 359 359 + .map(|(k, v)| Ok(encode(v).map(|v| (k.to_bytes(), v))?)); 360 360 + writer.put_many(kvs)?; 361 361 + } 362 362 + 363 363 + writer.commit()?; 364 364 + Ok::<_, DriveError>(store) 365 365 + }); // await later 206 366 207 367 // dump the rest to disk (in chunks) 368 368 + log::debug!("dumping the rest of the stream..."); 208 369 loop { 370 370 + let mut mem_size = 0; 209 371 let mut chunk = vec![]; 210 372 loop { 211 373 let Some((cid, data)) = self.car.next_block().await? else { ··· 219 381 } 220 382 // remaining possible types: node, record, other. optimistically process 221 383 // TODO: get the actual in-memory size to compute disk spill 222 222 - let maybe_processed = if Node::could_be(&data) { 223 223 - MaybeProcessedBlock::Raw(data) 224 224 - } else { 225 225 - MaybeProcessedBlock::Processed((self.process)(&data)) 226 226 - }; 384 384 + let maybe_processed = MaybeProcessedBlock::maybe(self.process, data); 385 385 + mem_size += std::mem::size_of::<Cid>() + maybe_processed.get_size(); 227 386 chunk.push((cid, maybe_processed)); 228 228 - if chunk.len() >= self.max_size { 229 229 - // eventually this won't be .len() 387 387 + if mem_size >= self.max_size { 388 388 + // soooooo if we're setting the db cache to max_size and then letting 389 389 + // multiple chunks in the queue that are >= max_size, then at any time 390 390 + // we might be using some multiple of max_size? 230 391 break; 231 392 } 232 393 } 233 394 if chunk.is_empty() { 234 395 break; 235 396 } 397 397 + tx.send(chunk) 398 398 + .await 399 399 + .map_err(|_| DriveError::ChannelSendError)?; 400 400 + } 401 401 + drop(tx); 402 402 + log::debug!("done. waiting for worker to finish..."); 236 403 237 237 - // move access in and back out so we can manage lifetimes 238 238 - // dump mem blocks into the store 239 239 - access = tokio::task::spawn_blocking(move || { 240 240 - let mut writer = access.get_writer()?; 241 241 - for (k, v) in chunk { 242 242 - let key_bytes = k.to_bytes(); 243 243 - let val_bytes = encode(v)?; // TODO 244 244 - writer.put(key_bytes, val_bytes)?; 245 245 - } 246 246 - drop(writer); // cannot outlive access 247 247 - Ok::<_, DiskDriveError<S::StorageError>>(access) 248 248 - }) 249 249 - .await 250 250 - .unwrap()?; // TODO 251 251 - } 404 404 + store = store_worker.await??; 405 405 + 406 406 + log::debug!("worker finished."); 252 407 253 253 - let commit = self.commit.ok_or(DiskDriveError::MissingCommit)?; 408 408 + let commit = self.commit.ok_or(DriveError::MissingCommit)?; 254 409 255 410 let walker = Walker::new(commit.data); 256 411 257 412 Ok(( 258 413 commit, 259 259 - BigCarReady { 414 414 + DiskDriver { 260 415 process: self.process, 261 261 - access, 262 262 - walker, 416 416 + state: Some(BigState { store, walker }), 263 417 }, 264 418 )) 265 419 } 266 420 } 267 421 268 268 - pub struct BigCarReady<T: Clone, A: DiskAccess> { 269 269 - process: fn(&[u8]) -> T, 270 270 - access: A, 422 422 + struct BigState { 423 423 + store: DiskStore, 271 424 walker: Walker, 272 425 } 273 426 274 274 - impl<T: Processable + Send + 'static, A: DiskAccess + Send + 'static> BigCarReady<T, A> { 275 275 - pub async fn next_chunk( 276 276 - mut self, 427 427 + /// MST walker that reads from disk instead of an in-memory hashmap 428 428 + pub struct DiskDriver<T: Clone> { 429 429 + process: fn(Vec<u8>) -> T, 430 430 + state: Option<BigState>, 431 431 + } 432 432 + 433 433 + // for doctests only 434 434 + #[doc(hidden)] 435 435 + pub fn _get_fake_disk_driver() -> DiskDriver<Vec<u8>> { 436 436 + use crate::process::noop; 437 437 + DiskDriver { 438 438 + process: noop, 439 439 + state: None, 440 440 + } 441 441 + } 442 442 + 443 443 + impl<T: Processable + Send + 'static> DiskDriver<T> { 444 444 + /// Walk the MST returning up to `n` rkey + record pairs 445 445 + /// 446 446 + /// ```no_run 447 447 + /// # use repo_stream::{drive::{DiskDriver, DriveError, _get_fake_disk_driver}, process::noop}; 448 448 + /// # #[tokio::main] 449 449 + /// # async fn main() -> Result<(), DriveError> { 450 450 + /// # let mut disk_driver = _get_fake_disk_driver(); 451 451 + /// while let Some(pairs) = disk_driver.next_chunk(256).await? { 452 452 + /// for (rkey, record) in pairs { 453 453 + /// println!("{rkey}: size={}", record.len()); 454 454 + /// } 455 455 + /// } 456 456 + /// let store = disk_driver.reset_store().await?; 457 457 + /// # Ok(()) 458 458 + /// # } 459 459 + /// ``` 460 460 + pub async fn next_chunk(&mut self, n: usize) -> Result<Option<BlockChunk<T>>, DriveError> { 461 461 + let process = self.process; 462 462 + 463 463 + // state should only *ever* be None transiently while inside here 464 464 + let mut state = self.state.take().expect("DiskDriver must have Some(state)"); 465 465 + 466 466 + // the big pain here is that we don't want to leave self.state in an 467 467 + // invalid state (None), so all the error paths have to make sure it 468 468 + // comes out again. 469 469 + let (state, res) = tokio::task::spawn_blocking( 470 470 + move || -> (BigState, Result<BlockChunk<T>, DriveError>) { 471 471 + let mut reader_res = state.store.get_reader(); 472 472 + let reader: &mut _ = match reader_res { 473 473 + Ok(ref mut r) => r, 474 474 + Err(ref mut e) => { 475 475 + // unfortunately we can't return the error directly because 476 476 + // (for some reason) it's attached to the lifetime of the 477 477 + // reader? 478 478 + // hack a mem::swap so we can get it out :/ 479 479 + let e_swapped = e.steal(); 480 480 + // the pain: `state` *has to* outlive the reader 481 481 + drop(reader_res); 482 482 + return (state, Err(e_swapped.into())); 483 483 + } 484 484 + }; 485 485 + 486 486 + let mut out = Vec::with_capacity(n); 487 487 + 488 488 + for _ in 0..n { 489 489 + // walk as far as we can until we run out of blocks or find a record 490 490 + let step = match state.walker.disk_step(reader, process) { 491 491 + Ok(s) => s, 492 492 + Err(e) => { 493 493 + // the pain: `state` *has to* outlive the reader 494 494 + drop(reader_res); 495 495 + return (state, Err(e.into())); 496 496 + } 497 497 + }; 498 498 + match step { 499 499 + Step::Missing(cid) => { 500 500 + // the pain: `state` *has to* outlive the reader 501 501 + drop(reader_res); 502 502 + return (state, Err(DriveError::MissingBlock(cid))); 503 503 + } 504 504 + Step::Finish => break, 505 505 + Step::Found { rkey, data } => out.push((rkey, data)), 506 506 + }; 507 507 + } 508 508 + 509 509 + // `state` *has to* outlive the reader 510 510 + drop(reader_res); 511 511 + 512 512 + (state, Ok::<_, DriveError>(out)) 513 513 + }, 514 514 + ) 515 515 + .await?; // on tokio JoinError, we'll be left with invalid state :( 516 516 + 517 517 + // *must* restore state before dealing with the actual result 518 518 + self.state = Some(state); 519 519 + 520 520 + let out = res?; 521 521 + 522 522 + if out.is_empty() { 523 523 + Ok(None) 524 524 + } else { 525 525 + Ok(Some(out)) 526 526 + } 527 527 + } 528 528 + 529 529 + fn read_tx_blocking( 530 530 + &mut self, 277 531 n: usize, 278 278 - ) -> Result<(Self, Option<Vec<(String, T)>>), DiskDriveError<A::StorageError>> 279 279 - where 280 280 - A::StorageError: Send, 281 281 - { 282 282 - let mut out = Vec::with_capacity(n); 283 283 - (self, out) = tokio::task::spawn_blocking(move || { 284 284 - let access = self.access; 285 285 - let mut reader = access.get_reader()?; 532 532 + tx: mpsc::Sender<Result<BlockChunk<T>, DriveError>>, 533 533 + ) -> Result<(), mpsc::error::SendError<Result<BlockChunk<T>, DriveError>>> { 534 534 + let BigState { store, walker } = self.state.as_mut().expect("valid state"); 535 535 + let mut reader = match store.get_reader() { 536 536 + Ok(r) => r, 537 537 + Err(e) => return tx.blocking_send(Err(e.into())), 538 538 + }; 539 539 + 540 540 + loop { 541 541 + let mut out: BlockChunk<T> = Vec::with_capacity(n); 286 542 287 543 for _ in 0..n { 288 544 // walk as far as we can until we run out of blocks or find a record 289 289 - match self.walker.disk_step(&mut reader, self.process)? { 290 290 - Step::Missing(cid) => return Err(DiskDriveError::MissingBlock(cid)), 291 291 - Step::Finish => break, 292 292 - Step::Step { rkey, data } => { 545 545 + 546 546 + let step = match walker.disk_step(&mut reader, self.process) { 547 547 + Ok(s) => s, 548 548 + Err(e) => return tx.blocking_send(Err(e.into())), 549 549 + }; 550 550 + 551 551 + match step { 552 552 + Step::Missing(cid) => { 553 553 + return tx.blocking_send(Err(DriveError::MissingBlock(cid))); 554 554 + } 555 555 + Step::Finish => return Ok(()), 556 556 + Step::Found { rkey, data } => { 293 557 out.push((rkey, data)); 294 558 continue; 295 559 } 296 560 }; 297 561 } 298 562 299 299 - drop(reader); // cannot outlive access 300 300 - self.access = access; 301 301 - Ok::<_, DiskDriveError<A::StorageError>>((self, out)) 302 302 - }) 303 303 - .await 304 304 - .unwrap()?; // TODO 563 563 + if out.is_empty() { 564 564 + break; 565 565 + } 566 566 + tx.blocking_send(Ok(out))?; 567 567 + } 305 568 306 306 - if out.is_empty() { 307 307 - Ok((self, None)) 308 308 - } else { 309 309 - Ok((self, Some(out))) 310 310 - } 569 569 + Ok(()) 311 570 } 312 312 - } 571 571 + 572 572 + /// Spawn the disk reading task into a tokio blocking thread 573 573 + /// 574 574 + /// The idea is to avoid so much sending back and forth to the blocking 575 575 + /// thread, letting a blocking task do all the disk reading work and sending 576 576 + /// records and rkeys back through an `mpsc` channel instead. 577 577 + /// 578 578 + /// This might also allow the disk work to continue while processing the 579 579 + /// records. It's still not yet clear if this method actually has much 580 580 + /// benefit over just using `.next_chunk(n)`. 581 581 + /// 582 582 + /// ```no_run 583 583 + /// # use repo_stream::{drive::{DiskDriver, DriveError, _get_fake_disk_driver}, process::noop}; 584 584 + /// # #[tokio::main] 585 585 + /// # async fn main() -> Result<(), DriveError> { 586 586 + /// # let mut disk_driver = _get_fake_disk_driver(); 587 587 + /// let (mut rx, join) = disk_driver.to_channel(512); 588 588 + /// while let Some(recvd) = rx.recv().await { 589 589 + /// let pairs = recvd?; 590 590 + /// for (rkey, record) in pairs { 591 591 + /// println!("{rkey}: size={}", record.len()); 592 592 + /// } 593 593 + /// 594 594 + /// } 595 595 + /// let store = join.await?.reset_store().await?; 596 596 + /// # Ok(()) 597 597 + /// # } 598 598 + /// ``` 599 599 + pub fn to_channel( 600 600 + mut self, 601 601 + n: usize, 602 602 + ) -> ( 603 603 + mpsc::Receiver<Result<BlockChunk<T>, DriveError>>, 604 604 + tokio::task::JoinHandle<Self>, 605 605 + ) { 606 606 + let (tx, rx) = mpsc::channel::<Result<BlockChunk<T>, DriveError>>(1); 313 607 314 314 - /// The core driver between the block stream and MST walker 315 315 - /// 316 316 - /// In the future, PDSs will export CARs in a stream-friendly order that will 317 317 - /// enable processing them with tiny memory overhead. But that future is not 318 318 - /// here yet. 319 319 - /// 320 320 - /// CARs are almost always in a stream-unfriendly order, so I'm reverting the 321 321 - /// optimistic stream features: we load all block first, then walk the MST. 322 322 - /// 323 323 - /// This makes things much simpler: we only need to worry about spilling to disk 324 324 - /// in one place, and we always have a reasonable expecatation about how much 325 325 - /// work the init function will do. We can drop the CAR reader before walking, 326 326 - /// so the sync/async boundaries become a little easier to work around. 327 327 - #[derive(Debug)] 328 328 - pub struct MemDriver<T: Processable> { 329 329 - blocks: HashMap<Cid, MaybeProcessedBlock<T>>, 330 330 - walker: Walker, 331 331 - process: fn(&[u8]) -> T, 332 332 - } 608 608 + // sketch: this worker is going to be allowed to execute without a join handle 609 609 + let chan_task = tokio::task::spawn_blocking(move || { 610 610 + if let Err(mpsc::error::SendError(_)) = self.read_tx_blocking(n, tx) { 611 611 + log::debug!("big car reader exited early due to dropped receiver channel"); 612 612 + } 613 613 + self 614 614 + }); 333 615 334 334 - impl<T: Processable> MemDriver<T> { 335 335 - /// Manually step through the record outputs 336 336 - pub async fn next_chunk(&mut self, n: usize) -> Result<Option<Vec<(String, T)>>, DriveError> { 337 337 - let mut out = Vec::with_capacity(n); 338 338 - for _ in 0..n { 339 339 - // walk as far as we can until we run out of blocks or find a record 340 340 - match self.walker.step(&mut self.blocks, self.process)? { 341 341 - Step::Missing(cid) => return Err(DriveError::MissingBlock(cid)), 342 342 - Step::Finish => break, 343 343 - Step::Step { rkey, data } => { 344 344 - out.push((rkey, data)); 345 345 - continue; 346 346 - } 347 347 - }; 348 348 - } 616 616 + (rx, chan_task) 617 617 + } 349 618 350 350 - if out.is_empty() { 351 351 - Ok(None) 352 352 - } else { 353 353 - Ok(Some(out)) 354 354 - } 619 619 + /// Reset the disk storage so it can be reused. You must call this. 620 620 + /// 621 621 + /// Ideally we'd put this in an `impl Drop`, but since it makes blocking 622 622 + /// calls, that would be risky in an async context. For now you just have to 623 623 + /// carefully make sure you call it. 624 624 + /// 625 625 + /// The sqlite store is returned, so it can be reused for another 626 626 + /// `DiskDriver`. 627 627 + pub async fn reset_store(mut self) -> Result<DiskStore, DriveError> { 628 628 + let BigState { store, .. } = self.state.take().expect("valid state"); 629 629 + Ok(store.reset().await?) 355 630 } 356 631 }

+84 -9

src/lib.rs

··· 1 1 - //! Fast and robust atproto CAR file processing in rust 2 2 - //! 3 3 - //! For now see the [examples](https://tangled.org/@microcosm.blue/repo-stream/tree/main/examples) 1 1 + /*! 2 2 + A robust CAR file -> MST walker for atproto 3 3 + 4 4 + Small CARs have their blocks buffered in memory. If a configurable memory limit 5 5 + is reached while reading blocks, CAR reading is suspended, and can be continued 6 6 + by providing disk storage to buffer the CAR blocks instead. 7 7 + 8 8 + A `process` function can be provided for tasks where records are transformed 9 9 + into a smaller representation, to save memory (and disk) during block reading. 10 10 + 11 11 + Once blocks are loaded, the MST is walked and emitted as chunks of pairs of 12 12 + `(rkey, processed_block)` pairs, in order (depth first, left-to-right). 13 13 + 14 14 + Some MST validations are applied 15 15 + - Keys must appear in order 16 16 + - Keys must be at the correct MST tree depth 17 17 + 18 18 + `iroh_car` additionally applies a block size limit of `2MiB`. 19 19 + 20 20 + ``` 21 21 + use repo_stream::{Driver, DriverBuilder, DiskBuilder}; 22 22 + 23 23 + # #[tokio::main] 24 24 + # async fn main() -> Result<(), Box<dyn std::error::Error>> { 25 25 + # let reader = include_bytes!("../car-samples/tiny.car").as_slice(); 26 26 + let mut total_size = 0; 27 27 + 28 28 + match DriverBuilder::new() 29 29 + .with_mem_limit_mb(10) 30 30 + .with_block_processor(|rec| rec.len()) // block processing: just extract the raw record size 31 31 + .load_car(reader) 32 32 + .await? 33 33 + { 34 34 + 35 35 + // if all blocks fit within memory 36 36 + Driver::Memory(_commit, mut driver) => { 37 37 + while let Some(chunk) = driver.next_chunk(256).await? { 38 38 + for (_rkey, size) in chunk { 39 39 + total_size += size; 40 40 + } 41 41 + } 42 42 + }, 43 43 + 44 44 + // if the CAR was too big for in-memory processing 45 45 + Driver::Disk(paused) => { 46 46 + // set up a disk store we can spill to 47 47 + let store = DiskBuilder::new().open("some/path.db".into()).await?; 48 48 + // do the spilling, get back a (similar) driver 49 49 + let (_commit, mut driver) = paused.finish_loading(store).await?; 50 50 + 51 51 + while let Some(chunk) = driver.next_chunk(256).await? { 52 52 + for (_rkey, size) in chunk { 53 53 + total_size += size; 54 54 + } 55 55 + } 56 56 + 57 57 + // clean up the disk store (drop tables etc) 58 58 + driver.reset_store().await?; 59 59 + } 60 60 + }; 61 61 + println!("sum of size of all records: {total_size}"); 62 62 + # Ok(()) 63 63 + # } 64 64 + ``` 65 65 + 66 66 + Disk spilling suspends and returns a `Driver::Disk(paused)` instead of going 67 67 + ahead and eagerly using disk I/O. This means you have to write a bit more code 68 68 + to handle both cases, but it allows you to have finer control over resource 69 69 + usage. For example, you can drive a number of parallel memory CAR workers, and 70 70 + separately have a different number of disk workers picking up suspended disk 71 71 + tasks from a queue. 72 72 + 73 73 + Find more [examples in the repo](https://tangled.org/@microcosm.blue/repo-stream/tree/main/examples). 74 74 + 75 75 + */ 76 76 + 77 77 + pub mod mst; 78 78 + mod walk; 4 79 5 80 pub mod disk; 6 6 - pub mod disk_drive; 7 7 - pub mod disk_redb; 8 8 - pub mod disk_sqlite; 9 9 - pub mod disk_walk; 10 81 pub mod drive; 11 11 - pub mod mst; 12 12 - pub mod walk; 82 82 + pub mod process; 83 83 + 84 84 + pub use disk::{DiskBuilder, DiskError, DiskStore}; 85 85 + pub use drive::{DriveError, Driver, DriverBuilder, NeedDisk}; 86 86 + pub use mst::Commit; 87 87 + pub use process::Processable;

+4 -8

src/mst.rs

··· 39 39 /// MST node data schema 40 40 #[derive(Debug, Deserialize, PartialEq)] 41 41 #[serde(deny_unknown_fields)] 42 42 - pub struct Node { 42 42 + pub(crate) struct Node { 43 43 /// link to sub-tree Node on a lower level and with all keys sorting before 44 44 /// keys at this node 45 45 #[serde(rename = "l")] ··· 62 62 /// so if a block *could be* a node, any record converter must postpone 63 63 /// processing. if it turns out it happens to be a very node-looking record, 64 64 /// well, sorry, it just has to only be processed later when that's known. 65 65 - pub fn could_be(bytes: impl AsRef<[u8]>) -> bool { 65 65 + pub(crate) fn could_be(bytes: impl AsRef<[u8]>) -> bool { 66 66 const NODE_FINGERPRINT: [u8; 3] = [ 67 67 0xA2, // map length 2 (for "l" and "e" keys) 68 68 0x61, // text length 1 ··· 83 83 /// with an empty array of entries. This is the only situation in which a 84 84 /// tree may contain an empty leaf node which does not either contain keys 85 85 /// ("entries") or point to a sub-tree containing entries. 86 86 - /// 87 87 - /// TODO: to me this is slightly unclear with respect to `l` (ask someone). 88 88 - /// ...is that what "The top of the tree must not be a an empty node which 89 89 - /// only points to a sub-tree." is referring to? 90 90 - pub fn is_empty(&self) -> bool { 86 86 + pub(crate) fn is_empty(&self) -> bool { 91 87 self.left.is_none() && self.entries.is_empty() 92 88 } 93 89 } ··· 95 91 /// TreeEntry object 96 92 #[derive(Debug, Deserialize, PartialEq)] 97 93 #[serde(deny_unknown_fields)] 98 98 - pub struct Entry { 94 94 + pub(crate) struct Entry { 99 95 /// count of bytes shared with previous TreeEntry in this Node (if any) 100 96 #[serde(rename = "p")] 101 97 pub prefix_len: usize,

+108

src/process.rs

··· 1 1 + /*! 2 2 + Record processor function output trait 3 3 + 4 4 + The return type must satisfy the `Processable` trait, which requires: 5 5 + 6 6 + - `Clone` because two rkeys can refer to the same record by CID, which may 7 7 + only appear once in the CAR file. 8 8 + - `Serialize + DeserializeOwned` so it can be spilled to disk. 9 9 + 10 10 + One required function must be implemented, `get_size()`: this should return the 11 11 + approximate total off-stack size of the type. (the on-stack size will be added 12 12 + automatically via `std::mem::get_size`). 13 13 + 14 14 + Note that it is **not guaranteed** that the `process` function will run on a 15 15 + block before storing it in memory or on disk: it's not possible to know if a 16 16 + block is a record without actually walking the MST, so the best we can do is 17 17 + apply `process` to any block that we know *cannot* be an MST node, and otherwise 18 18 + store the raw block bytes. 19 19 + 20 20 + Here's a silly processing function that just collects 'eyy's found in the raw 21 21 + record bytes 22 22 + 23 23 + ``` 24 24 + # use repo_stream::Processable; 25 25 + # use serde::{Serialize, Deserialize}; 26 26 + #[derive(Debug, Clone, Serialize, Deserialize)] 27 27 + struct Eyy(usize, String); 28 28 + 29 29 + impl Processable for Eyy { 30 30 + fn get_size(&self) -> usize { 31 31 + // don't need to compute the usize, it's on the stack 32 32 + self.1.capacity() // in-mem size from the string's capacity, in bytes 33 33 + } 34 34 + } 35 35 + 36 36 + fn process(raw: Vec<u8>) -> Vec<Eyy> { 37 37 + let mut out = Vec::new(); 38 38 + let to_find = "eyy".as_bytes(); 39 39 + for i in 0..(raw.len() - 3) { 40 40 + if &raw[i..(i+3)] == to_find { 41 41 + out.push(Eyy(i, "eyy".to_string())); 42 42 + } 43 43 + } 44 44 + out 45 45 + } 46 46 + ``` 47 47 + 48 48 + The memory sizing stuff is a little sketch but probably at least approximately 49 49 + works. 50 50 + */ 51 51 + 52 52 + use serde::{Serialize, de::DeserializeOwned}; 53 53 + 54 54 + /// Output trait for record processing 55 55 + pub trait Processable: Clone + Serialize + DeserializeOwned { 56 56 + /// Any additional in-memory size taken by the processed type 57 57 + /// 58 58 + /// Do not include stack size (`std::mem::size_of`) 59 59 + fn get_size(&self) -> usize; 60 60 + } 61 61 + 62 62 + /// Processor that just returns the raw blocks 63 63 + #[inline] 64 64 + pub fn noop(block: Vec<u8>) -> Vec<u8> { 65 65 + block 66 66 + } 67 67 + 68 68 + impl Processable for u8 { 69 69 + fn get_size(&self) -> usize { 70 70 + 0 71 71 + } 72 72 + } 73 73 + 74 74 + impl Processable for usize { 75 75 + fn get_size(&self) -> usize { 76 76 + 0 // no additional space taken, just its stack size (newtype is free) 77 77 + } 78 78 + } 79 79 + 80 80 + impl Processable for String { 81 81 + fn get_size(&self) -> usize { 82 82 + self.capacity() 83 83 + } 84 84 + } 85 85 + 86 86 + impl<Item: Sized + Processable> Processable for Vec<Item> { 87 87 + fn get_size(&self) -> usize { 88 88 + let slot_size = std::mem::size_of::<Item>(); 89 89 + let direct_size = slot_size * self.capacity(); 90 90 + let items_referenced_size: usize = self.iter().map(|item| item.get_size()).sum(); 91 91 + direct_size + items_referenced_size 92 92 + } 93 93 + } 94 94 + 95 95 + impl<Item: Processable> Processable for Option<Item> { 96 96 + fn get_size(&self) -> usize { 97 97 + self.as_ref().map(|item| item.get_size()).unwrap_or(0) 98 98 + } 99 99 + } 100 100 + 101 101 + impl<Item: Processable, Error: Processable> Processable for Result<Item, Error> { 102 102 + fn get_size(&self) -> usize { 103 103 + match self { 104 104 + Ok(item) => item.get_size(), 105 105 + Err(err) => err.get_size(), 106 106 + } 107 107 + } 108 108 + }

+201 -266

src/walk.rs

··· 1 1 //! Depth-first MST traversal 2 2 3 3 - use crate::disk::{DiskReader, StorageErrorBase}; 4 4 - use crate::drive::{MaybeProcessedBlock, Processable}; 3 3 + use crate::disk::SqliteReader; 4 4 + use crate::drive::{DecodeError, MaybeProcessedBlock}; 5 5 use crate::mst::Node; 6 6 + use crate::process::Processable; 6 7 use ipld_core::cid::Cid; 8 8 + use sha2::{Digest, Sha256}; 7 9 use std::collections::HashMap; 8 10 use std::convert::Infallible; 9 11 10 12 /// Errors that can happen while walking 11 13 #[derive(Debug, thiserror::Error)] 12 12 - pub enum Trip { 13 13 - #[error("empty mst nodes are not allowed")] 14 14 - NodeEmpty, 14 14 + pub enum WalkError { 15 15 #[error("Failed to fingerprint commit block")] 16 16 BadCommitFingerprint, 17 17 #[error("Failed to decode commit block: {0}")] 18 18 BadCommit(#[from] serde_ipld_dagcbor::DecodeError<Infallible>), 19 19 #[error("Action node error: {0}")] 20 20 - RkeyError(#[from] RkeyError), 21 21 - #[error("Encountered an rkey out of order while walking the MST")] 22 22 - RkeyOutOfOrder, 23 23 - } 24 24 - 25 25 - /// Errors that can happen while walking 26 26 - #[derive(Debug, thiserror::Error)] 27 27 - pub enum DiskTrip<E: StorageErrorBase> { 28 28 - #[error("tripped: {0}")] 29 29 - Trip(#[from] Trip), 20 20 + MstError(#[from] MstError), 30 21 #[error("storage error: {0}")] 31 31 - StorageError(#[from] E), 22 22 + StorageError(#[from] rusqlite::Error), 32 23 #[error("Decode error: {0}")] 33 33 - BincodeDecodeError(#[from] bincode::error::DecodeError), 24 24 + DecodeError(#[from] DecodeError), 34 25 } 35 26 36 27 /// Errors from invalid Rkeys 37 37 - #[derive(Debug, thiserror::Error)] 38 38 - pub enum RkeyError { 28 28 + #[derive(Debug, PartialEq, thiserror::Error)] 29 29 + pub enum MstError { 39 30 #[error("Failed to compute an rkey due to invalid prefix_len")] 40 31 EntryPrefixOutOfbounds, 41 32 #[error("RKey was not utf-8")] 42 33 EntryRkeyNotUtf8(#[from] std::string::FromUtf8Error), 34 34 + #[error("Nodes cannot be empty (except for an entirely empty MST)")] 35 35 + EmptyNode, 36 36 + #[error("Found an entry with rkey at the wrong depth")] 37 37 + WrongDepth, 38 38 + #[error("Lost track of our depth (possible bug?)")] 39 39 + LostDepth, 40 40 + #[error("MST depth underflow: depth-0 node with child trees")] 41 41 + DepthUnderflow, 42 42 + #[error("Encountered an rkey out of order while walking the MST")] 43 43 + RkeyOutOfOrder, 43 44 } 44 45 45 46 /// Walker outputs ··· 50 51 /// Reached the end of the MST! yay! 51 52 Finish, 52 53 /// A record was found! 53 53 - Step { rkey: String, data: T }, 54 54 + Found { rkey: String, data: T }, 54 55 } 55 56 56 57 #[derive(Debug, Clone, PartialEq)] 57 58 enum Need { 58 58 - Node(Cid), 59 59 + Node { depth: Depth, cid: Cid }, 59 60 Record { rkey: String, cid: Cid }, 60 61 } 61 62 62 62 - fn push_from_node(stack: &mut Vec<Need>, node: &Node) -> Result<(), RkeyError> { 63 63 + #[derive(Debug, Clone, Copy, PartialEq)] 64 64 + enum Depth { 65 65 + Root, 66 66 + Depth(u32), 67 67 + } 68 68 + 69 69 + impl Depth { 70 70 + fn from_key(key: &[u8]) -> Self { 71 71 + let mut zeros = 0; 72 72 + for byte in Sha256::digest(key) { 73 73 + let leading = byte.leading_zeros(); 74 74 + zeros += leading; 75 75 + if leading < 8 { 76 76 + break; 77 77 + } 78 78 + } 79 79 + Self::Depth(zeros / 2) // truncating divide (rounds down) 80 80 + } 81 81 + fn next_expected(&self) -> Result<Option<u32>, MstError> { 82 82 + match self { 83 83 + Self::Root => Ok(None), 84 84 + Self::Depth(d) => d.checked_sub(1).ok_or(MstError::DepthUnderflow).map(Some), 85 85 + } 86 86 + } 87 87 + } 88 88 + 89 89 + fn push_from_node(stack: &mut Vec<Need>, node: &Node, parent_depth: Depth) -> Result<(), MstError> { 90 90 + // empty nodes are not allowed in the MST except in an empty MST 91 91 + if node.is_empty() { 92 92 + if parent_depth == Depth::Root { 93 93 + return Ok(()); // empty mst, nothing to push 94 94 + } else { 95 95 + return Err(MstError::EmptyNode); 96 96 + } 97 97 + } 98 98 + 63 99 let mut entries = Vec::with_capacity(node.entries.len()); 64 64 - 65 100 let mut prefix = vec![]; 101 101 + let mut this_depth = parent_depth.next_expected()?; 102 102 + 66 103 for entry in &node.entries { 67 104 let mut rkey = vec![]; 68 105 let pre_checked = prefix 69 106 .get(..entry.prefix_len) 70 70 - .ok_or(RkeyError::EntryPrefixOutOfbounds)?; 107 107 + .ok_or(MstError::EntryPrefixOutOfbounds)?; 71 108 rkey.extend_from_slice(pre_checked); 72 109 rkey.extend_from_slice(&entry.keysuffix); 110 110 + 111 111 + let Depth::Depth(key_depth) = Depth::from_key(&rkey) else { 112 112 + return Err(MstError::WrongDepth); 113 113 + }; 114 114 + 115 115 + // this_depth is `none` if we are the deepest child (directly below root) 116 116 + // in that case we accept whatever highest depth is claimed 117 117 + let expected_depth = match this_depth { 118 118 + Some(d) => d, 119 119 + None => { 120 120 + this_depth = Some(key_depth); 121 121 + key_depth 122 122 + } 123 123 + }; 124 124 + 125 125 + // all keys we find should be this depth 126 126 + if key_depth != expected_depth { 127 127 + return Err(MstError::DepthUnderflow); 128 128 + } 129 129 + 73 130 prefix = rkey.clone(); 74 131 75 132 entries.push(Need::Record { ··· 77 134 cid: entry.value, 78 135 }); 79 136 if let Some(ref tree) = entry.tree { 80 80 - entries.push(Need::Node(*tree)); 137 137 + entries.push(Need::Node { 138 138 + depth: Depth::Depth(key_depth), 139 139 + cid: *tree, 140 140 + }); 81 141 } 82 142 } 83 143 84 144 entries.reverse(); 85 145 stack.append(&mut entries); 86 146 147 147 + let d = this_depth.ok_or(MstError::LostDepth)?; 148 148 + 87 149 if let Some(tree) = node.left { 88 88 - stack.push(Need::Node(tree)); 150 150 + stack.push(Need::Node { 151 151 + depth: Depth::Depth(d), 152 152 + cid: tree, 153 153 + }); 89 154 } 90 155 Ok(()) 91 156 } ··· 102 167 impl Walker { 103 168 pub fn new(tree_root_cid: Cid) -> Self { 104 169 Self { 105 105 - stack: vec![Need::Node(tree_root_cid)], 170 170 + stack: vec![Need::Node { 171 171 + depth: Depth::Root, 172 172 + cid: tree_root_cid, 173 173 + }], 106 174 prev: "".to_string(), 107 175 } 108 176 } ··· 111 179 pub fn step<T: Processable>( 112 180 &mut self, 113 181 blocks: &mut HashMap<Cid, MaybeProcessedBlock<T>>, 114 114 - process: impl Fn(&[u8]) -> T, 115 115 - ) -> Result<Step<T>, Trip> { 182 182 + process: impl Fn(Vec<u8>) -> T, 183 183 + ) -> Result<Step<T>, WalkError> { 116 184 loop { 117 117 - let Some(mut need) = self.stack.last() else { 185 185 + let Some(need) = self.stack.last_mut() else { 118 186 log::trace!("tried to walk but we're actually done."); 119 187 return Ok(Step::Finish); 120 188 }; 121 189 122 122 - match &mut need { 123 123 - Need::Node(cid) => { 190 190 + match need { 191 191 + &mut Need::Node { depth, cid } => { 124 192 log::trace!("need node {cid:?}"); 125 125 - let Some(block) = blocks.remove(cid) else { 193 193 + let Some(block) = blocks.remove(&cid) else { 126 194 log::trace!("node not found, resting"); 127 127 - return Ok(Step::Missing(*cid)); 195 195 + return Ok(Step::Missing(cid)); 128 196 }; 129 197 130 198 let MaybeProcessedBlock::Raw(data) = block else { 131 131 - return Err(Trip::BadCommitFingerprint); 199 199 + return Err(WalkError::BadCommitFingerprint); 132 200 }; 133 133 - let node = 134 134 - serde_ipld_dagcbor::from_slice::<Node>(&data).map_err(Trip::BadCommit)?; 201 201 + let node = serde_ipld_dagcbor::from_slice::<Node>(&data) 202 202 + .map_err(WalkError::BadCommit)?; 135 203 136 204 // found node, make sure we remember 137 205 self.stack.pop(); 138 206 139 207 // queue up work on the found node next 140 140 - push_from_node(&mut self.stack, &node)?; 208 208 + push_from_node(&mut self.stack, &node, depth)?; 141 209 } 142 210 Need::Record { rkey, cid } => { 143 211 log::trace!("need record {cid:?}"); 212 212 + // note that we cannot *remove* a record block, sadly, since 213 213 + // there can be multiple rkeys pointing to the same cid. 144 214 let Some(data) = blocks.get_mut(cid) else { 145 145 - log::trace!("record block not found, resting"); 146 215 return Ok(Step::Missing(*cid)); 147 216 }; 148 217 let rkey = rkey.clone(); 149 218 let data = match data { 150 150 - MaybeProcessedBlock::Raw(data) => process(data), 219 219 + MaybeProcessedBlock::Raw(data) => process(data.to_vec()), 151 220 MaybeProcessedBlock::Processed(t) => t.clone(), 152 221 }; 153 222 154 223 // found node, make sure we remember 155 224 self.stack.pop(); 156 156 - 157 157 - log::trace!("emitting a block as a step. depth={}", self.stack.len()); 158 225 159 226 // rkeys *must* be in order or else the tree is invalid (or 160 227 // we have a bug) 161 228 if rkey <= self.prev { 162 162 - return Err(Trip::RkeyOutOfOrder); 229 229 + return Err(MstError::RkeyOutOfOrder)?; 163 230 } 164 231 self.prev = rkey.clone(); 165 232 166 166 - return Ok(Step::Step { rkey, data }); 233 233 + return Ok(Step::Found { rkey, data }); 167 234 } 168 235 } 169 236 } 170 237 } 171 238 172 239 /// blocking!!!!!! 173 173 - pub fn disk_step<T: Processable, R: DiskReader>( 240 240 + pub fn disk_step<T: Processable>( 174 241 &mut self, 175 175 - reader: &mut R, 176 176 - process: impl Fn(&[u8]) -> T, 177 177 - ) -> Result<Step<T>, DiskTrip<R::StorageError>> { 242 242 + reader: &mut SqliteReader, 243 243 + process: impl Fn(Vec<u8>) -> T, 244 244 + ) -> Result<Step<T>, WalkError> { 178 245 loop { 179 179 - let Some(mut need) = self.stack.last() else { 246 246 + let Some(need) = self.stack.last_mut() else { 180 247 log::trace!("tried to walk but we're actually done."); 181 248 return Ok(Step::Finish); 182 249 }; 183 250 184 184 - match &mut need { 185 185 - Need::Node(cid) => { 251 251 + match need { 252 252 + &mut Need::Node { depth, cid } => { 186 253 let cid_bytes = cid.to_bytes(); 187 254 log::trace!("need node {cid:?}"); 188 255 let Some(block_bytes) = reader.get(cid_bytes)? else { 189 256 log::trace!("node not found, resting"); 190 190 - return Ok(Step::Missing(*cid)); 257 257 + return Ok(Step::Missing(cid)); 191 258 }; 192 259 193 260 let block: MaybeProcessedBlock<T> = crate::drive::decode(&block_bytes)?; 194 261 195 262 let MaybeProcessedBlock::Raw(data) = block else { 196 196 - return Err(Trip::BadCommitFingerprint.into()); 263 263 + return Err(WalkError::BadCommitFingerprint); 197 264 }; 198 198 - let node = 199 199 - serde_ipld_dagcbor::from_slice::<Node>(&data).map_err(Trip::BadCommit)?; 265 265 + let node = serde_ipld_dagcbor::from_slice::<Node>(&data) 266 266 + .map_err(WalkError::BadCommit)?; 200 267 201 268 // found node, make sure we remember 202 269 self.stack.pop(); 203 270 204 271 // queue up work on the found node next 205 205 - push_from_node(&mut self.stack, &node).map_err(Trip::RkeyError)?; 272 272 + push_from_node(&mut self.stack, &node, depth).map_err(WalkError::MstError)?; 206 273 } 207 274 Need::Record { rkey, cid } => { 208 275 log::trace!("need record {cid:?}"); ··· 214 281 let data: MaybeProcessedBlock<T> = crate::drive::decode(&data_bytes)?; 215 282 let rkey = rkey.clone(); 216 283 let data = match data { 217 217 - MaybeProcessedBlock::Raw(data) => process(&data), 284 284 + MaybeProcessedBlock::Raw(data) => process(data), 218 285 MaybeProcessedBlock::Processed(t) => t.clone(), 219 286 }; 220 287 ··· 226 293 // rkeys *must* be in order or else the tree is invalid (or 227 294 // we have a bug) 228 295 if rkey <= self.prev { 229 229 - return Err(DiskTrip::Trip(Trip::RkeyOutOfOrder)); 296 296 + return Err(MstError::RkeyOutOfOrder)?; 230 297 } 231 298 self.prev = rkey.clone(); 232 299 233 233 - return Ok(Step::Step { rkey, data }); 300 300 + return Ok(Step::Found { rkey, data }); 234 301 } 235 302 } 236 303 } ··· 240 307 #[cfg(test)] 241 308 mod test { 242 309 use super::*; 243 243 - // use crate::mst::Entry; 244 310 245 311 fn cid1() -> Cid { 246 312 "bafyreihixenvk3ahqbytas4hk4a26w43bh6eo3w6usjqtxkpzsvi655a3m" 247 313 .parse() 248 314 .unwrap() 249 315 } 250 250 - // fn cid2() -> Cid { 251 251 - // "QmY7Yh4UquoXHLPFo2XbhXkhBvFoPwmQUSa92pxnxjQuPU" 252 252 - // .parse() 253 253 - // .unwrap() 254 254 - // } 255 255 - // fn cid3() -> Cid { 256 256 - // "bafybeigdyrzt5sfp7udm7hu76uh7y26nf3efuylqabf3oclgtqy55fbzdi" 257 257 - // .parse() 258 258 - // .unwrap() 259 259 - // } 260 260 - // fn cid4() -> Cid { 261 261 - // "QmbWqxBEKC3P8tqsKc98xmWNzrzDtRLMiMPL8wBuTGsMnR" 262 262 - // .parse() 263 263 - // .unwrap() 264 264 - // } 265 265 - // fn cid5() -> Cid { 266 266 - // "QmSnuWmxptJZdLJpKRarxBMS2Ju2oANVrgbr2xWbie9b2D" 267 267 - // .parse() 268 268 - // .unwrap() 269 269 - // } 270 270 - // fn cid6() -> Cid { 271 271 - // "QmdmQXB2mzChmMeKY47C43LxUdg1NDJ5MWcKMKxDu7RgQm" 272 272 - // .parse() 273 273 - // .unwrap() 274 274 - // } 275 275 - // fn cid7() -> Cid { 276 276 - // "bafybeiaysi4s6lnjev27ln5icwm6tueaw2vdykrtjkwiphwekaywqhcjze" 277 277 - // .parse() 278 278 - // .unwrap() 279 279 - // } 280 280 - // fn cid8() -> Cid { 281 281 - // "bafyreif3tfdpr5n4jdrbielmcapwvbpcthepfkwq2vwonmlhirbjmotedi" 282 282 - // .parse() 283 283 - // .unwrap() 284 284 - // } 285 285 - // fn cid9() -> Cid { 286 286 - // "bafyreicnokmhmrnlp2wjhyk2haep4tqxiptwfrp2rrs7rzq7uk766chqvq" 287 287 - // .parse() 288 288 - // .unwrap() 289 289 - // } 316 316 + 317 317 + #[test] 318 318 + fn test_depth_spec_0() { 319 319 + let d = Depth::from_key(b"2653ae71"); 320 320 + assert_eq!(d, Depth::Depth(0)) 321 321 + } 322 322 + 323 323 + #[test] 324 324 + fn test_depth_spec_1() { 325 325 + let d = Depth::from_key(b"blue"); 326 326 + assert_eq!(d, Depth::Depth(1)) 327 327 + } 328 328 + 329 329 + #[test] 330 330 + fn test_depth_spec_4() { 331 331 + let d = Depth::from_key(b"app.bsky.feed.post/454397e440ec"); 332 332 + assert_eq!(d, Depth::Depth(4)) 333 333 + } 334 334 + 335 335 + #[test] 336 336 + fn test_depth_spec_8() { 337 337 + let d = Depth::from_key(b"app.bsky.feed.post/9adeb165882c"); 338 338 + assert_eq!(d, Depth::Depth(8)) 339 339 + } 340 340 + 341 341 + #[test] 342 342 + fn test_depth_ietf_draft_0() { 343 343 + let d = Depth::from_key(b"key1"); 344 344 + assert_eq!(d, Depth::Depth(0)) 345 345 + } 346 346 + 347 347 + #[test] 348 348 + fn test_depth_ietf_draft_1() { 349 349 + let d = Depth::from_key(b"key7"); 350 350 + assert_eq!(d, Depth::Depth(1)) 351 351 + } 352 352 + 353 353 + #[test] 354 354 + fn test_depth_ietf_draft_4() { 355 355 + let d = Depth::from_key(b"key515"); 356 356 + assert_eq!(d, Depth::Depth(4)) 357 357 + } 358 358 + 359 359 + #[test] 360 360 + fn test_depth_interop() { 361 361 + // examples from https://github.com/bluesky-social/atproto-interop-tests/blob/main/mst/key_heights.json 362 362 + for (k, expected) in [ 363 363 + ("", 0), 364 364 + ("asdf", 0), 365 365 + ("blue", 1), 366 366 + ("2653ae71", 0), 367 367 + ("88bfafc7", 2), 368 368 + ("2a92d355", 4), 369 369 + ("884976f5", 6), 370 370 + ("app.bsky.feed.post/454397e440ec", 4), 371 371 + ("app.bsky.feed.post/9adeb165882c", 8), 372 372 + ] { 373 373 + let d = Depth::from_key(k.as_bytes()); 374 374 + assert_eq!(d, Depth::Depth(expected), "key: {}", k); 375 375 + } 376 376 + } 290 377 291 378 #[test] 292 292 - fn test_next_from_node_empty() { 293 293 - let node = Node { 379 379 + fn test_push_empty_fails() { 380 380 + let empty_node = Node { 294 381 left: None, 295 382 entries: vec![], 296 383 }; 297 384 let mut stack = vec![]; 298 298 - push_from_node(&mut stack, &node).unwrap(); 299 299 - assert_eq!(stack.last(), None); 385 385 + let err = push_from_node(&mut stack, &empty_node, Depth::Depth(4)); 386 386 + assert_eq!(err, Err(MstError::EmptyNode)); 300 387 } 301 388 302 389 #[test] 303 303 - fn test_needs_from_node_just_left() { 390 390 + fn test_push_one_node() { 304 391 let node = Node { 305 392 left: Some(cid1()), 306 393 entries: vec![], 307 394 }; 308 395 let mut stack = vec![]; 309 309 - push_from_node(&mut stack, &node).unwrap(); 310 310 - assert_eq!(stack.last(), Some(Need::Node(cid1())).as_ref()); 396 396 + push_from_node(&mut stack, &node, Depth::Depth(4)).unwrap(); 397 397 + assert_eq!( 398 398 + stack.last(), 399 399 + Some(Need::Node { 400 400 + depth: Depth::Depth(3), 401 401 + cid: cid1() 402 402 + }) 403 403 + .as_ref() 404 404 + ); 311 405 } 312 312 - 313 313 - // #[test] 314 314 - // fn test_needs_from_node_just_one_record() { 315 315 - // let node = Node { 316 316 - // left: None, 317 317 - // entries: vec![Entry { 318 318 - // keysuffix: "asdf".into(), 319 319 - // prefix_len: 0, 320 320 - // value: cid1(), 321 321 - // tree: None, 322 322 - // }], 323 323 - // }; 324 324 - // assert_eq!( 325 325 - // needs_from_node(node).unwrap(), 326 326 - // vec![Need::Record { 327 327 - // rkey: "asdf".into(), 328 328 - // cid: cid1(), 329 329 - // },] 330 330 - // ); 331 331 - // } 332 332 - 333 333 - // #[test] 334 334 - // fn test_needs_from_node_two_records() { 335 335 - // let node = Node { 336 336 - // left: None, 337 337 - // entries: vec![ 338 338 - // Entry { 339 339 - // keysuffix: "asdf".into(), 340 340 - // prefix_len: 0, 341 341 - // value: cid1(), 342 342 - // tree: None, 343 343 - // }, 344 344 - // Entry { 345 345 - // keysuffix: "gh".into(), 346 346 - // prefix_len: 2, 347 347 - // value: cid2(), 348 348 - // tree: None, 349 349 - // }, 350 350 - // ], 351 351 - // }; 352 352 - // assert_eq!( 353 353 - // needs_from_node(node).unwrap(), 354 354 - // vec![ 355 355 - // Need::Record { 356 356 - // rkey: "asdf".into(), 357 357 - // cid: cid1(), 358 358 - // }, 359 359 - // Need::Record { 360 360 - // rkey: "asgh".into(), 361 361 - // cid: cid2(), 362 362 - // }, 363 363 - // ] 364 364 - // ); 365 365 - // } 366 366 - 367 367 - // #[test] 368 368 - // fn test_needs_from_node_with_both() { 369 369 - // let node = Node { 370 370 - // left: None, 371 371 - // entries: vec![Entry { 372 372 - // keysuffix: "asdf".into(), 373 373 - // prefix_len: 0, 374 374 - // value: cid1(), 375 375 - // tree: Some(cid2()), 376 376 - // }], 377 377 - // }; 378 378 - // assert_eq!( 379 379 - // needs_from_node(node).unwrap(), 380 380 - // vec![ 381 381 - // Need::Record { 382 382 - // rkey: "asdf".into(), 383 383 - // cid: cid1(), 384 384 - // }, 385 385 - // Need::Node(cid2()), 386 386 - // ] 387 387 - // ); 388 388 - // } 389 389 - 390 390 - // #[test] 391 391 - // fn test_needs_from_node_left_and_record() { 392 392 - // let node = Node { 393 393 - // left: Some(cid1()), 394 394 - // entries: vec![Entry { 395 395 - // keysuffix: "asdf".into(), 396 396 - // prefix_len: 0, 397 397 - // value: cid2(), 398 398 - // tree: None, 399 399 - // }], 400 400 - // }; 401 401 - // assert_eq!( 402 402 - // needs_from_node(node).unwrap(), 403 403 - // vec![ 404 404 - // Need::Node(cid1()), 405 405 - // Need::Record { 406 406 - // rkey: "asdf".into(), 407 407 - // cid: cid2(), 408 408 - // }, 409 409 - // ] 410 410 - // ); 411 411 - // } 412 412 - 413 413 - // #[test] 414 414 - // fn test_needs_from_full_node() { 415 415 - // let node = Node { 416 416 - // left: Some(cid1()), 417 417 - // entries: vec![ 418 418 - // Entry { 419 419 - // keysuffix: "asdf".into(), 420 420 - // prefix_len: 0, 421 421 - // value: cid2(), 422 422 - // tree: Some(cid3()), 423 423 - // }, 424 424 - // Entry { 425 425 - // keysuffix: "ghi".into(), 426 426 - // prefix_len: 1, 427 427 - // value: cid4(), 428 428 - // tree: Some(cid5()), 429 429 - // }, 430 430 - // Entry { 431 431 - // keysuffix: "jkl".into(), 432 432 - // prefix_len: 2, 433 433 - // value: cid6(), 434 434 - // tree: Some(cid7()), 435 435 - // }, 436 436 - // Entry { 437 437 - // keysuffix: "mno".into(), 438 438 - // prefix_len: 4, 439 439 - // value: cid8(), 440 440 - // tree: Some(cid9()), 441 441 - // }, 442 442 - // ], 443 443 - // }; 444 444 - // assert_eq!( 445 445 - // needs_from_node(node).unwrap(), 446 446 - // vec![ 447 447 - // Need::Node(cid1()), 448 448 - // Need::Record { 449 449 - // rkey: "asdf".into(), 450 450 - // cid: cid2(), 451 451 - // }, 452 452 - // Need::Node(cid3()), 453 453 - // Need::Record { 454 454 - // rkey: "aghi".into(), 455 455 - // cid: cid4(), 456 456 - // }, 457 457 - // Need::Node(cid5()), 458 458 - // Need::Record { 459 459 - // rkey: "agjkl".into(), 460 460 - // cid: cid6(), 461 461 - // }, 462 462 - // Need::Node(cid7()), 463 463 - // Need::Record { 464 464 - // rkey: "agjkmno".into(), 465 465 - // cid: cid8(), 466 466 - // }, 467 467 - // Need::Node(cid9()), 468 468 - // ] 469 469 - // ); 470 470 - // } 471 406 }

+34 -31

tests/non-huge-cars.rs

··· 1 1 extern crate repo_stream; 2 2 - use futures::TryStreamExt; 3 3 - use iroh_car::CarReader; 4 4 - use std::convert::Infallible; 2 2 + use repo_stream::Driver; 5 3 4 4 + const EMPTY_CAR: &'static [u8] = include_bytes!("../car-samples/empty.car"); 6 5 const TINY_CAR: &'static [u8] = include_bytes!("../car-samples/tiny.car"); 7 6 const LITTLE_CAR: &'static [u8] = include_bytes!("../car-samples/little.car"); 8 7 const MIDSIZE_CAR: &'static [u8] = include_bytes!("../car-samples/midsize.car"); 9 8 10 10 - async fn test_car(bytes: &[u8], expected_records: usize, expected_sum: usize) { 11 11 - let reader = CarReader::new(bytes).await.unwrap(); 12 12 - 13 13 - let root = reader 14 14 - .header() 15 15 - .roots() 16 16 - .first() 17 17 - .ok_or("missing root") 9 9 + async fn test_car( 10 10 + bytes: &[u8], 11 11 + expected_records: usize, 12 12 + expected_sum: usize, 13 13 + expect_profile: bool, 14 14 + ) { 15 15 + let mut driver = match Driver::load_car(bytes, |block| block.len(), 10 /* MiB */) 16 16 + .await 18 17 .unwrap() 19 19 - .clone(); 20 20 - 21 21 - let stream = std::pin::pin!(reader.stream()); 22 22 - 23 23 - let (_commit, v) = 24 24 - repo_stream::drive::Vehicle::init(root, stream, |block| Ok::<_, Infallible>(block.len())) 25 25 - .await 26 26 - .unwrap(); 27 27 - let mut record_stream = std::pin::pin!(v.stream()); 18 18 + { 19 19 + Driver::Memory(_commit, mem_driver) => mem_driver, 20 20 + Driver::Disk(_) => panic!("too big"), 21 21 + }; 28 22 29 23 let mut records = 0; 30 24 let mut sum = 0; 31 25 let mut found_bsky_profile = false; 32 26 let mut prev_rkey = "".to_string(); 33 33 - while let Some((rkey, size)) = record_stream.try_next().await.unwrap() { 34 34 - records += 1; 35 35 - sum += size; 36 36 - if rkey == "app.bsky.actor.profile/self" { 37 37 - found_bsky_profile = true; 27 27 + 28 28 + while let Some(pairs) = driver.next_chunk(256).await.unwrap() { 29 29 + for (rkey, size) in pairs { 30 30 + records += 1; 31 31 + sum += size; 32 32 + if rkey == "app.bsky.actor.profile/self" { 33 33 + found_bsky_profile = true; 34 34 + } 35 35 + assert!(rkey > prev_rkey, "rkeys are streamed in order"); 36 36 + prev_rkey = rkey; 38 37 } 39 39 - assert!(rkey > prev_rkey, "rkeys are streamed in order"); 40 40 - prev_rkey = rkey; 41 38 } 39 39 + 42 40 assert_eq!(records, expected_records); 43 41 assert_eq!(sum, expected_sum); 44 44 - assert!(found_bsky_profile); 42 42 + assert_eq!(found_bsky_profile, expect_profile); 43 43 + } 44 44 + 45 45 + #[tokio::test] 46 46 + async fn test_empty_car() { 47 47 + test_car(EMPTY_CAR, 0, 0, false).await 45 48 } 46 49 47 50 #[tokio::test] 48 51 async fn test_tiny_car() { 49 49 - test_car(TINY_CAR, 8, 2071).await 52 52 + test_car(TINY_CAR, 8, 2071, true).await 50 53 } 51 54 52 55 #[tokio::test] 53 56 async fn test_little_car() { 54 54 - test_car(LITTLE_CAR, 278, 246960).await 57 57 + test_car(LITTLE_CAR, 278, 246960, true).await 55 58 } 56 59 57 60 #[tokio::test] 58 61 async fn test_midsize_car() { 59 59 - test_car(MIDSIZE_CAR, 11585, 3741393).await 62 62 + test_car(MIDSIZE_CAR, 11585, 3741393, true).await 60 63 }