Summary:
We removed the feature in D20704618 and it does not cause complaints.
Let's remove the code supporting the chown feature.
Reviewed By: DurhamG
Differential Revision: D21170307
fbshipit-source-id: c845016219e8c681930bb1780b94e6d31ca99730
Summary:
RotateLog loads older logs lazily. If an older log is broken, remember that and avoid
loading the broken log again.
Reviewed By: DurhamG
Differential Revision: D20663899
fbshipit-source-id: 7a4b5279cc6387c19329a51048bfe1be2e0bc1f8
Summary:
These were from a wide variety of warnings. The only one I haven't addressed is
that clippy complains that Pin<Box<Vec<u8>>> can be replaced by Pin<Vec<u8>>. I
haven't investigated too much into it, someone more familiar with this code can
probably figure out if this is buggy or not :)
Reviewed By: DurhamG
Differential Revision: D20469647
fbshipit-source-id: d42891d95c1d21b625230234994ab49bbc45b961
Summary:
This belongs to D20149376. However buck test does not include benchmarks so it
was not noticed.
Reviewed By: DurhamG
Differential Revision: D20505097
fbshipit-source-id: 24daeb17b68808f8e69e18452ab2cf26c7aa10a7
Summary:
Some `use`s are not used on Windows. The code was also formatted using the
latest rustfmt.
Reviewed By: xavierd
Differential Revision: D20379704
fbshipit-source-id: ffadcd68e4e0440dcbd2a4e1ad8532b47a9d83e2
Summary:
Recently there are some Windows-related test flakiness in . All of them are
caused by `file.persist(path)` in `atomic_write_plain` failing with
"Access Denied". Since that can be caused by Windows Anti-Virus scans or other
weird stuff, let's workaround around it using automatically retires.
Process Explorer does not provide extra information:
indexedlog-d0c6135fd7ed9ece.exe 5868 SetRenameInformationFile C:\Users\quark\AppData\Local\Temp\.tmpKERc5G\.tmpcfDsQQ ACCESS DENIED ReplaceIfExists: True, FileName: C:\Users\quark\AppData\Local\Temp\.tmpKERc5G\meta
A successful rename looks like:
indexedlog-d0c6135fd7ed9ece.exe 5868 SetRenameInformationFile C:\Users\quark\AppData\Local\Temp\.tmpKERc5G\.tmpbXEVw0 SUCCESS ReplaceIfExists: True, FileName: C:\Users\quark\AppData\Local\Temp\.tmpKERc5G\meta
Reviewed By: ikostia
Differential Revision: D20379618
fbshipit-source-id: db3e6be3d785875486f7a517df11cbf58bf65ddd
Summary: This makes `RUST_LOG` work for indexedlog tests.
Reviewed By: xavierd
Differential Revision: D20286515
fbshipit-source-id: ff4a1476eb01a9067dabe3622fd598f65fe86a18
Summary: Those has helped me debugging some issues.
Reviewed By: xavierd
Differential Revision: D20286513
fbshipit-source-id: 012ddb16c2d0efd8f8697a5ecd4564ea31d65630
Summary:
The old code does "read, lock, write", which is unsound because after "lock"
the data just read can be outdated and needs a reload.
Reviewed By: xavierd
Differential Revision: D20306137
fbshipit-source-id: a1c29d5078b2d47ee95cf00db8c1fcbe3447cccf
Summary:
I thought the index function could be the bottleneck. However, the Log reading
(xxhash, decoding vlqs) can be much slower for very long entries. Therefore
using bytes as the lag threshold is better. It does leaked the Log
implementation details (how it encodes an entry) to some extend, though.
Reverts D20042045 and D20043116 logically. The lagging calculation is using
the new Index::get_original_meta API, which is easier to verify correctness
(In fact, it seems the old code is wrong - it might skip Index flushes if
sync() is called multiple times without flushing).
This should mitigate an issue where a huge entry (generated by `hg trace`) in
blackbox does not get indexed in time and cause performance regressions.
Reviewed By: DurhamG
Differential Revision: D20286508
fbshipit-source-id: 7cd694b58b95537490047fb1834c16b30d102f18
Summary: This will be used to more reliably detect index lags.
Reviewed By: DurhamG
Differential Revision: D20286518
fbshipit-source-id: c553b6587363a55603b75df12580588e3100e35f
Summary:
This ensures indexes are complete even if index format or definition has been
changed.
Reviewed By: DurhamG
Differential Revision: D20286509
fbshipit-source-id: fcc4ebc616a4501e4b6fd2f1a9826f54f40b99b8
Summary: This API has the benefit that it does not trigger loading older logs.
Reviewed By: DurhamG
Differential Revision: D20286512
fbshipit-source-id: 426421691ad1130cdbb2305612d76f18c9f8798c
Summary:
Previously, `flush()` will skip writing the file if there are only metadata
changes. Fix it by detecting metadata changes.
This can potentially fix an issue that certain blackbox indexes are empty,
lagging and require scanning the whole log again and again. In that case,
the index itself is not changed (the root radix entry is not changed), but
only the metadata tracking how many bytes in Log the index covered
changed.
Reviewed By: sfilipco
Differential Revision: D20264627
fbshipit-source-id: 7ee48454a92b5786b847d8b1d738cc38183f7a32
Summary:
The change is in theory not necessary. However it improves the reliability on
OS crashes a bit, and can potentially workaround some bugs in filesystems
(as we saw in production where the atomic-written files are empty and the
system didn't crash).
The idea is, the `symlink` syscall does the file creation and "content" writing
together, while there is no way to create a file and write specific content
in one syscall. Note that the C symlink call uses 0-terminated string, and
the Rust stdlib exports it as accepting `Path`. To be safe, we encode binary
or non-utf8 content using `hex`.
For downgrade safety, the write path does not use symlink by default unless
format.use-symlink-atomic-write is set to true. This makes downgrade possible:
the read path is rolled out first, then we can turn on and off the write path.
The indexedlog Rust unit tests and test-doctor.t are migrated to use the new
symlink code paths.
Reviewed By: DurhamG
Differential Revision: D20153864
fbshipit-source-id: c31bd4287a8d29575180fbcf7227d2b04c4c1252
Summary:
This makes it possible to implement atomic_write differently (ex. use a
symlink).
Reviewed By: DurhamG
Differential Revision: D20153865
fbshipit-source-id: 07fa78c2f2dac696668f477c75f65cf70950b73f
Summary:
This makes it clear that `log` is a math concept, not an append-only file like
`Log`.
Reviewed By: DurhamG
Differential Revision: D20149376
fbshipit-source-id: 67d2e9584b15f48759ca9b6dfce4279a5b1365a0
Summary:
Previously indexes are only updated at `sync()` time. This diff makes it so
`open()` can also update lagging indexes. This should make index migration
(ex. D19851355) smoother - indexes are built in time and users suffer less from
the absent of indexes.
Reviewed By: DurhamG
Differential Revision: D20042046
fbshipit-source-id: 20412661a0ca4f5f67b671137c47b6373a42981d
Summary: The logic is currently only used by `sync()`. I'd like to reuse it at `open()`.
Reviewed By: DurhamG
Differential Revision: D20042044
fbshipit-source-id: 5c9734ff68bdcf8f8c8710c6a821b18d3afeaca0
Summary:
This is more friendly for indexedlog users - deciding lag_threshold by number
of entries is easier than by bytes.
Initially, I thought checking `bytes` is cheaper and checking `entries` is more
expensive. However, practically we will have to build indexes for `entires`
anyway. So we do know the number of entries lagging behind.
Reviewed By: DurhamG
Differential Revision: D20042045
fbshipit-source-id: 73042e406bd8b262d5ef9875e45a3fd5f29f78cf
Summary:
This can be useful for users of indexedlog when they want `Bytes` (to get rid
of the lifetime parameter).
This might be useful for storage layer that wants to take the ownership of the
returned bytes.
Reviewed By: xavierd
Differential Revision: D19818714
fbshipit-source-id: cb2d4e7deff921915e07454fee15cb94a3d5c00d
Summary: Those utilities are no longer necessary since the new code uses Bytes.
Reviewed By: xavierd
Differential Revision: D19818717
fbshipit-source-id: 0b43af0f1eae1a4288e84d4170db058b27f80334
Summary: This simplifies the code a bit and makes it cheaper to clone the Log.
Reviewed By: xavierd
Differential Revision: D19818716
fbshipit-source-id: bbf07b8b36009d53b63d8066ec422fc3c3796840
Summary: It's no longer used since Index now has inlined its checksum logic.
Reviewed By: ikostia
Differential Revision: D19850744
fbshipit-source-id: eb134e4c1613573a2d238710b44ad8119c80a5ee
Summary:
Change index filename and metadata name. This makes sure the new format and old
format are separate so upgrading or downgrading won't have issues.
Reviewed By: DurhamG
Differential Revision: D19851355
fbshipit-source-id: 25dee018073a90040f5818b32b753a3f589c10e0
Summary:
Enhance the index format: The Root entry can be followed by an optional
Checksum entry which replaces the need of ChecksumTable.
The format is backwards compatible since the old format will be just
treated as "there is no ChecksumTable", and the ChecksumTable will be built on
the next "flush".
This change is non-trivial. But the tests are pretty strong - the bitflip test
alone covered a lot of issues, and the dump of Index content helps a lot too.
For the index itself without ".sum", checksum, this change is bi-directional
compatible:
1. New code reading old file will just think the old file does not have the
checksum entry, similar to new code having checksum disabled.
2. Old code will think the root+checksum slice is the "root" entry. Parsing
the root entry is fine since it does not complain about unknown data at the
end.
However, this change dropped the logic updating ".sum" files. That part is an
issue blocking old clients from reading new data.
Reviewed By: DurhamG
Differential Revision: D19850741
fbshipit-source-id: 551a45cd5422f1fb4c5b08e3b207a2ffe3d93dea
Summary:
To solve the soundness issue of ChecksumTable raised by the last diff.
I plan to move Checksum logic to Index. This has multiple benefits:
- Solve the soundness issue of ChecksumTable.
- Indexedlog no longer writes the ".sum" files. `atomic_write` can be quite
slow (tens of milliseconds) on Windows. So this should help perf - with
many indexes, it can save hundreds of milliseconds on Windows per
indexedlog sync.
This diff adds the definition and serialization of the new Checksum entry.
The index format is not updated yet.
Reviewed By: markbt
Differential Revision: D19850742
fbshipit-source-id: df6e6ed12a12ef0d2a782dc9d6b4dc5dec3f4b46
Summary:
With the last change, mmap cost is reduced, but ChecksumTable is unsound in a
corner case: the buffer to check is shorter than what ChecksumTable covers:
checksum: |----chunk----|----chunk----|----chunk--|
buf: |-------------------------------| |
^ ^
logic len physical len
The checksum table will be unable to verify the last chunk, since it does not
have enough data in buf.
The issues is exposed by stress testing the multithread sync tests. It's not
always easy to reproduce, though.
Reviewed By: markbt
Differential Revision: D19850745
fbshipit-source-id: a1a96080163b7b9b56dcd6c1673d5d8d10e18a2b
Summary: This avoids some extra mmap syscalls by ChecksumTable.
Reviewed By: xavierd
Differential Revision: D19818721
fbshipit-source-id: dace55193f2b4b0f35e3868781faa2d2998d3b58
Summary:
This simplifies the code a bit (no special cases about 0-sized mmap buffers)
and makes it cheaper to clone the index buffer (just an Arc::clone, without
another mmap syscall).
Reviewed By: xavierd
Differential Revision: D19818718
fbshipit-source-id: e96d42af74c7f0bb11703c5da31cdfbd5d76c372
Summary:
This makes it possible to use `Bytes` for mmap buffers.
The changes are because `minibytes::Bytes` does not implement `From<&[u8]>`
with the intention to make slice copy explicit.
Reviewed By: xavierd
Differential Revision: D19818719
fbshipit-source-id: c34ee451bfd2dc7bcbbcebd52a76444b6c236849
Summary: This makes it possible to do extra sanity checks.
Reviewed By: DurhamG
Differential Revision: D19443783
fbshipit-source-id: 254c2537a6aadd25a67c5e48a768187ce65aa686
Summary: This makes the code overall shorter.
Reviewed By: DurhamG
Differential Revision: D19443552
fbshipit-source-id: abd1db227830a88549c7eca1cfd08b67c4914518
Summary:
In both cases (clone with or without dirty content), the external key buffers
used by indexes should be re-created, since mem_buf location has changed.
Reviewed By: DurhamG
Differential Revision: D19432657
fbshipit-source-id: fe6f76e7ccfd16ccd2f5c1d89866687a3503603e
Summary: This allows us to build the multi-log structure.
Reviewed By: DurhamG
Differential Revision: D19431786
fbshipit-source-id: e0e09b0d5a73d293a80626924b74ddf2ce6d6fa3
Summary: Migrate to the new API. This is more compatible with the next change.
Reviewed By: DurhamG
Differential Revision: D19431788
fbshipit-source-id: dba1a88fd003d17fc1774b0cec9157d32c5acdb0
Summary:
This allows us to incrementally abstract away parts related to filesystem from
Log. For example, instead of using std::fs to access filesystem directly, use
methods on the GenericPath instead.
Right now the main motivation is to add support for "multi-logs" sharing a single meta
file. To do that, methods reading and writing metas are moved to the
GenericPath type.
This also unifies the open functions. Now OpenOptions::create_in_memory can be
private. Empty in-memory Logs can be opened via `open(())`.
Reviewed By: DurhamG
Differential Revision: D19431784
fbshipit-source-id: dbdf94f60261e09f131c6fdd9fe3b99242a28af5
Summary: This makes an upcoming change easier.
Reviewed By: DurhamG
Differential Revision: D19432656
fbshipit-source-id: 65adc883c3c3937aa7022196b83c75715b820721
Summary: The goal is to split log.rs to smaller modules so it's easier to reason about.
Reviewed By: DurhamG
Differential Revision: D19431787
fbshipit-source-id: 23b8346c461da322e7a2240b8224e6194867aaee
Summary:
The goal is to split log.rs to smaller modules so it's easier to reason about.
OpenOptions has many dependencies. Move the most complicated one (repair) to
another standalone module.
Reviewed By: DurhamG
Differential Revision: D19431783
fbshipit-source-id: 28c3af7cd6e19588bfe9c395e2648244faf3179d
Summary: The goal is to split log.rs to smaller modules so it's easier to reason about.
Reviewed By: DurhamG
Differential Revision: D19431785
fbshipit-source-id: b9a598b3018267ff6e43ef57dc807cb4e880467c
Summary:
Previously, between 100 to 500 runs, the test is likely to fail:
---- log::tests::test_repair_and_delete_content stdout ----
thread 'log::tests::test_repair_and_delete_content' panicked at 'assertion failed: `(left == right)`
left: `"Verified first 141 entries, 999 of 1400012 bytes in log\nReset log size to 999\nIndex \"c\" is incompatible with (truncated) log\nRebuilt index \"c\""`,
right: `"Rebuilt metadata\nVerified first 141 entries, 999 of 1400012 bytes in log\nReset log size to 999\nRebuilt index \"c\""`', src/log.rs:3414:9
The "Rebuilt metadata" part is missing.
This is because after D17742003, the end of `meta` is the `epoch`, which is a
random number. The test is rewriting the end of `meta` to corrupt it. That
has a chance to not corrupt the file. Change the test to corrupt the fixed
header of meta so it can always corrupt the file and therefore stabilize
the test.
Reviewed By: xavierd
Differential Revision: D19325087
fbshipit-source-id: 53d46c8a6cb9771582f8f37e27f185d303fde0ce
Summary: Those warnings are not seen on POSIX build.
Reviewed By: sfilipco
Differential Revision: D19320441
fbshipit-source-id: feddf728eb9627834559b87d83e20f0afd9080c8
Summary:
The index can delete all keys matching a prefix more efficiently than deleting
them one by one. Expose this feature.
The `dag` crate will use this feature to delete all "non-master" segments and
ids efficiently.
Reviewed By: sfilipco
Differential Revision: D18825296
fbshipit-source-id: b8531695609238a16913254af61004170f12954e
Summary: This makes it possible to define an index function that can delete data.
Reviewed By: sfilipco
Differential Revision: D18825299
fbshipit-source-id: 4eb03a001683092d99b439c8efb8e7909f543c70
Summary:
This makes it possible to actually delete keys from the index, which can be
handy for certain use-cases. For example, `dag` wants to rewrite non-master
ids and segments in certain cases.
Reviewed By: sfilipco
Differential Revision: D18825298
fbshipit-source-id: 3b69ca5fc763ba0c57bd16a28a349f2f829a0759
