Summary:
Previously for point lookup we decided which file to look into based on user key overlap only. We also did not truncate range tombstones in the point lookup code path. These two ideas did not interact well in cases like this:
- L1 has range tombstone [a, c)#1 and point key b#2. The data is split between file1 with range [a#1,1, b#72057594037927935,15], and file2 with range [b#2, c#1].
- L1's file2 gets compacted to L2.
- User issues `Get()` for b#3.
- L1's file1 is opened and the range tombstone [a, c)#1 is found for b, while no point-key for b is found in L1.
- `Get()` assumes that the range tombstone must cover all data in that range in lower levels, so short circuits and returns `NotFound`.
The solution to this problem is to not look into files that only overlap with the point lookup at a range tombstone sentinel endpoint. In the above example, this would mean not opening L1's file1 or its tombstones during the `Get()`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4829
Differential Revision: D13561355
Pulled By: ajkr
fbshipit-source-id: a13c21c816870a2f5d32a48af6dbd719a7d9d19f
Summary:
When there's a gap between files, we do not need to output tombstones starting at the next output file's begin key to the current output file.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4592
Differential Revision: D12808627
Pulled By: ajkr
fbshipit-source-id: 77c8b2e7523a95b1cd6611194144092c06acb505
Summary:
When there are no range deletions, flush and compaction perform a binary search
on an effectively empty map every time they call ShouldDelete. This PR lazily
initializes each stripe map entry so that the binary search can be elided in
these cases.
After this PR, the total amount of time spent in compactions is 52.541331s, and the total amount of time spent in flush is 5.532608s, the former of which is a significant improvement from the results after #4495.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4497
Differential Revision: D10428610
Pulled By: abhimadan
fbshipit-source-id: 6f7e1ce3698fac3ef86d1197955e6b72e0931a0f
Summary:
Even during `DBIter::Prev()`, there is a case where we need to use `RangeDelPositioningMode::kForwardTraversal`. In particular, when we hit too many internal keys for a single user key, we use seek to find the newest internal key. If it's a merge operand, we then scan forwards, collecting the merge operands. This forward scan should be using `RangeDelPositioningMode::kForwardTraversal`.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4481
Differential Revision: D10319507
Pulled By: ajkr
fbshipit-source-id: b5ce7352461f3a7696b28a5136ae0076f2bde51f
Summary:
I wrote a couple tests using the public API to expose/prevent the bugs we talked. In particular,
- When files have overlapping endpoints and a range tombstone spans them, ensure the largest key does not reappear to readers. This was happening due to a bug that skipped writing range tombstones to an output file when their begin key exactly matched the file's largest key.
- When a tombstone spans multiple atomic compaction units, ensure newer keys do not disappear by being compacted beneath it. This happened due to a range tombstone appearing untruncated to readers when it spanned files with overlapping endpoints, even if it extended into files without overlapping endpoints (i.e., different atomic compaction units).
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4476
Differential Revision: D10286001
Pulled By: ajkr
fbshipit-source-id: bb5ca51d0f90812fb37bfe1d01aec93f7eda55aa
Summary:
RocksDB used to store global_seqno in external SST files written by
SstFileWriter. During file ingestion, RocksDB uses `pwrite` to update the
`global_seqno`. Since random write is not supported in some non-POSIX compliant
file systems, external SST file ingestion is not supported on these file
systems. To address this limitation, we no longer update `global_seqno` during
file ingestion. Later RocksDB uses the MANIFEST and other information in table
properties to deduce global seqno for externally-ingested SST files.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4172
Differential Revision: D8961465
Pulled By: riversand963
fbshipit-source-id: 4382ec85270a96be5bc0cf33758ca2b167b05071
Summary:
Fixes#3391.
This change adds a `DeleteRange` method to `SstFileWriter` and adds
support for ingesting SSTs with range deletion tombstones. This is
important for applications that need to atomically ingest SSTs while
clearing out any existing keys in a given key range.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/3778
Differential Revision: D8821836
Pulled By: anand1976
fbshipit-source-id: ca7786c1947ff129afa703dab011d524c7883844
Summary:
Do not consider the range tombstone sentinel key as causing 2 adjacent
sstables in a level to overlap. When a range tombstone's end key is the
largest key in an sstable, the sstable's end key is so to a "sentinel"
value that is the smallest key in the next sstable with a sequence
number of kMaxSequenceNumber. This "sentinel" is guaranteed to not
overlap in internal-key space with the next sstable. Unfortunately,
GetOverlappingFiles uses user-keys to determine overlap and was thus
considering 2 adjacent sstables in a level to overlap if they were
separated by this sentinel key. This in turn would cause compactions to
be larger than necessary.
Note that this conflicts with
https://github.com/facebook/rocksdb/pull/2769 and cases
`DBRangeDelTest.CompactionTreatsSplitInputLevelDeletionAtomically` to
fail.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4050
Differential Revision: D8844423
Pulled By: ajkr
fbshipit-source-id: df3f9f1db8f4cff2bff77376b98b83c2ae1d155b
Summary:
This fixes the same performance issue that #3992 fixes but with much more invasive cleanup.
I'm more excited about this PR because it paves the way for fixing another problem we uncovered at Cockroach where range deletion tombstones can cause massive compactions. For example, suppose L4 contains deletions from [a, c) and [x, z) and no other keys, and L5 is entirely empty. L6, however, is full of data. When compacting L4 -> L5, we'll end up with one file that spans, massively, from [a, z). When we go to compact L5 -> L6, we'll have to rewrite all of L6! If, instead of range deletions in L4, we had keys a, b, x, y, and z, RocksDB would have been smart enough to create two files in L5: one for a and b and another for x, y, and z.
With the changes in this PR, it will be possible to adjust the compaction logic to split tombstones/start new output files when they would span too many files in the grandparent level.
ajkr please take a look when you have a minute!
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4014
Differential Revision: D8773253
Pulled By: ajkr
fbshipit-source-id: ec62fa85f648fdebe1380b83ed997f9baec35677
Summary:
Run the basic range deletion tests against the standard set of
configurations. This testing exposed that files with hash indexes and
partitioned indexes were not handling the case where the file contained
only range deletions--i.e., where the index was empty.
Additionally file a TODO about the fact that range deletions are broken
when allow_mmap_reads = true is set.
/cc ajkr nvanbenschoten
Best viewed with ?w=1: https://github.com/facebook/rocksdb/pull/4021/files?w=1
Pull Request resolved: https://github.com/facebook/rocksdb/pull/4021
Differential Revision: D8811860
Pulled By: ajkr
fbshipit-source-id: 3cc07e6d6210a2a00b932866481b3d5c59775343
Summary:
Sometimes we want to compact files as fast as possible, but don't want to set a large `max_subcompactions` in the `DBOptions` by default.
I add a `max_subcompactions` options to `CompactionOptions` so that we can choose a proper concurrency dynamically.
Closes https://github.com/facebook/rocksdb/pull/3775
Differential Revision: D7792357
Pulled By: ajkr
fbshipit-source-id: 94f54c3784dce69e40a229721a79a97e80cd6a6c
Summary:
One of the core assumptions of DeleteRange is that files containing portions of the same range tombstone are treated as a single unit from the perspective of compaction picker. Need better tests for this. This PR adds the tests for manual compaction.
Closes https://github.com/facebook/rocksdb/pull/2769
Differential Revision: D5676677
Pulled By: ajkr
fbshipit-source-id: 1b4b3382b300ff7048b872911405fdf900e4fbec
Summary:
the range delete tombstones in memtable should be added to the aggregator even when the memtable's prefix bloom filter tells us the lookup key's not there. This bug could cause data to temporarily reappear until the memtable containing range deletions is flushed.
Reported in #2743.
Closes https://github.com/facebook/rocksdb/pull/2745
Differential Revision: D5639007
Pulled By: ajkr
fbshipit-source-id: 04fc6facb6f978340a3f639536f4ca7c0d73dfc9
Summary:
Issue: #2478Fix: #2503
The bug happened when all of these conditions were satisfied:
- A subcompaction generates no keys
- `RangeDelAggregator::ShouldAddTombstones()` returns true because there's at least one non-obsoleted range deletion in its map
- None of the non-obsolete tombstones overlap with the subcompaction key-range
Under those conditions, we were creating a dedicated file for range deletions which was left empty, thus causing an error in VersionEdit.
I verified this test case fails before the #2503 fix and passes after.
Closes https://github.com/facebook/rocksdb/pull/2521
Differential Revision: D5352568
Pulled By: ajkr
fbshipit-source-id: f619cae39984ce9bb9b7a4e7a9ac0f2bb2ce43e9
Summary:
This PR adds support for encrypting data stored by RocksDB when written to disk.
It adds an `EncryptedEnv` override of the `Env` class with matching overrides for sequential&random access files.
The encryption itself is done through a configurable `EncryptionProvider`. This class creates is asked to create `BlockAccessCipherStream` for a file. This is where the actual encryption/decryption is being done.
Currently there is a Counter mode implementation of `BlockAccessCipherStream` with a `ROT13` block cipher (NOTE the `ROT13` is for demo purposes only!!).
The Counter operation mode uses an initial counter & random initialization vector (IV).
Both are created randomly for each file and stored in a 4K (default size) block that is prefixed to that file. The `EncryptedEnv` implementation is such that clients of the `Env` class do not see this prefix (nor data, nor in filesize).
The largest part of the prefix block is also encrypted, and there is room left for implementation specific settings/values/keys in there.
To test the encryption, the `DBTestBase` class has been extended to consider a new environment variable called `ENCRYPTED_ENV`. If set, the test will setup a encrypted instance of the `Env` class to use for all tests.
Typically you would run it like this:
```
ENCRYPTED_ENV=1 make check_some
```
There is also an added test that checks that some data inserted into the database is or is not "visible" on disk. With `ENCRYPTED_ENV` active it must not find plain text strings, with `ENCRYPTED_ENV` unset, it must find the plain text strings.
Closes https://github.com/facebook/rocksdb/pull/2424
Differential Revision: D5322178
Pulled By: sdwilsh
fbshipit-source-id: 253b0a9c2c498cc98f580df7f2623cbf7678a27f
Summary:
UBSAN crashes when it run the test. Disabling it for UBSAN.
Closes https://github.com/facebook/rocksdb/pull/2427
Differential Revision: D5210897
Pulled By: yiwu-arbug
fbshipit-source-id: 2f5a876807c98d8db79ab9581965f7e6b29d4163
Summary:
Level-based L0->L0 compaction operates on spans of files that aren't currently being compacted. It reduces the number of L0 files, thus making write stall conditions harder to reach.
- L0->L0 is triggered when base level is unavailable due to pending compactions
- L0->L0 always outputs one file of at most `max_level0_burst_file_size` bytes.
- Subcompactions are disabled for L0->L0 since we want to output one file.
- Input files are chosen as the longest span of available files that will fit within the size limit. This minimizes number of files in L0.
Closes https://github.com/facebook/rocksdb/pull/2027
Differential Revision: D4760318
Pulled By: ajkr
fbshipit-source-id: 9d07183
Summary:
when writing RangeDelAggregator::AddToBuilder, I forgot that there are sentinel tombstones in the middle of the interval map since gaps between real tombstones are represented with sentinels.
blame: #1614
Closes https://github.com/facebook/rocksdb/pull/1804
Differential Revision: D4460426
Pulled By: ajkr
fbshipit-source-id: 69444b5
Summary:
It's a test case for #1797. Also got rid of kTypeDeletion in the conditional since we treat it the same as kTypeRangeDeletion.
Closes https://github.com/facebook/rocksdb/pull/1800
Differential Revision: D4451300
Pulled By: ajkr
fbshipit-source-id: b39dda1
Summary:
This test ensures RangeDelAggregator can still access blocks even if it outlives the table readers that created them (detailed description in comments).
I plan to optimize away the extra cache lookup we currently do in BlockBasedTable::NewRangeTombstoneIterator(), as it is ~5% CPU in my random read benchmark in a database with 1k tombstones. This test will help make sure nothing breaks in the process.
Closes https://github.com/facebook/rocksdb/pull/1739
Differential Revision: D4375954
Pulled By: ajkr
fbshipit-source-id: aef9357
Summary:
change the iterator status to NotSupported as soon as a range tombstone
is encountered by a ForwardIterator.
Closes https://github.com/facebook/rocksdb/pull/1593
Differential Revision: D4246294
Pulled By: ajkr
fbshipit-source-id: aef9f49
Summary:
Cockroachdb exposed this bug in #1778. The bug happens when a compaction's output files are ended due to exceeding max_compaction_bytes. In that case we weren't taking into account the next file's start key when deciding how far to extend the current file's max_key. This caused the non-overlapping key-range invariant to be violated.
Note this was correctly handled for the usual case of cutting compaction output, which is file size exceeding max_output_file_size. I am not sure why these are two separate code paths, but we can consider refactoring it to prevent such errors in the future.
Closes https://github.com/facebook/rocksdb/pull/1784
Differential Revision: D4430235
Pulled By: ajkr
fbshipit-source-id: 80af748
Summary:
Added a tombstone-collapsing mode to RangeDelAggregator, which eliminates overlap in the TombstoneMap. In this mode, we can check whether a tombstone covers a user key using upper_bound() (i.e., binary search). However, the tradeoff is the overhead to add tombstones is now higher, so at first I've only enabled it for range scans (compaction/flush/user iterators), where we expect a high number of calls to ShouldDelete() for the same tombstones. Point queries like Get() will still use the linear scan approach.
Also in this diff I changed RangeDelAggregator's TombstoneMap to use multimap with user keys instead of map with internal keys. Callers sometimes provided ParsedInternalKey directly, from which it would've required string copying to derive an internal key Slice with which we could search the map.
Closes https://github.com/facebook/rocksdb/pull/1614
Differential Revision: D4270397
Pulled By: ajkr
fbshipit-source-id: 93092c7
