feat: TopOrderable shares TopN threshold across segments - #65
Closed
rebasedming wants to merge 70 commits into
Closed
feat: TopOrderable shares TopN threshold across segments#65rebasedming wants to merge 70 commits into
TopOrderable shares TopN threshold across segments#65rebasedming wants to merge 70 commits into
Conversation
This PR modifies internal API signatures and implementation details so that `FileSlice`s are passed down into the innards of (at least) the `BlockwiseLinearCodec`. This allows tantivy to defer dereferencing large slices of bytes when reading numeric fast fields, and instead dereference only the slice of bytes it needs for any given compressed Block. The motivation here is for external `Directory` implementations where it's not exactly efficient to dereference large slices of bytes.
…mponents (#20) This overhauls `SegmentReader` to put its various components behind `OnceLock`s such that they can be opened and read on their first use, as oppoed when a SegmentReader is constructed -- which is once for every segment when an Index is opened. This has a negative impact on some of Tantivy's expectations in that an existing SegementReader can still read from physical files that were deleted by a merge. This isn't true now that the segment's physical files aren't opened until needed. As such, I've `#[ignore]`'d six tests that expose this problem. From our (pg_search's) side of things, we don't really have physical files and don't need to rely on the filesystem/kernel to allow reading unlinked files that are still open. Overall, this cuts down a signficiant number of disk reads during pg_search's query planning. With my test data it goes from 808 individual reads totalling 999,799 bytes, to 18 reads totalling 814,514 bytes. This reduces the time it takes to plan a simple query from about 1.4ms to 0.436ms -- roughly a 3.2x improvement.
This removes up some overhead the profiler exposed. In the case I was testing, fast fields no longer shows up in the profile at all. I also renamed `BlockWithLength` to `BlockWithData`
Prior to this commit, the Footer tantivy serialized at the end of every file included a json blob that could be an arbitrary size. This changes the Footer to be exactly 24 bytes (6 u32s), making sure to keep the `crc` value. The other change we make here is to not actually read/validate the footer bytes when opening a file. From pg_search's perspective, this is quite unnecessary Postgres buffer cache I/O and increases index/segment opening overhead, which is something pg_search does often for each query. Two tests are ignored here as they test physical index files stored here in the repo that this change completely breaks.
This allows a `SegmentId` to be constructed from a `[u8; 16]` byte array. It also adds a `impl Default for SegementId`, which defaults to all nulls
This teaches tantivy how to "directly" delete a document in a segment.
Our use case from pg_search is that we already know the segment_id and doc_id so it's waaaaay more efficient for us to delete docs through our `ambulkdelete()` routine.
It avoids doing a search, and all the stuff around that, for each of our "ctid" terms that we want to delete.
Tantivy creates thread pools for some of its background work, specifically committing and merging. It's possible if one of the thread workers panics that rayon will simply abort the process. This is terrible from pg_search as that takes down the entire Postgres cluster. These changes allow a Directory to assign a panic handler that gets called in such cases. Which allows pg_search to gracefully rollback the current transaction, while presenting the panic message to the user.
Not all queries use positions and it's okay if we (from the perspective of `pg_search`, anyways) defer opening/loading them until they're first needed. This is probably completely wrong for a mmap-based Directory, but we (again, `pg_search`) decided long ago that we don't care about that use case. This saves a lot of disk I/O when an index has lots of segments and the query doesn't need positions. As a drive by, make sure a random Vec has enough space before pushing items to it. This showed up in the profiler, believe it or not.
…tionary`. (#55) When a fast fields string/bytes `Dictionary` is opened, we currently read the entire dictionary from `FileSlice` -> `OwnedBytes`... and then immediately wrap it back into a `FileSlice`. Switching to `Dictionary::open` preserves the `FileSlice`, such that only the portions of the `Dictionary` which are actually accessed are read from disk/buffers.
We would like to be able to lazily load `BitpackedCodec` columns (similar to what 020bdff did for `BlockwiseLinearCodec`), because in the context of `pg_search`, immediately constructing `OwnedBytes` means copying the entire content of the column into memory. To do so, we expose some (slightly overlapped) block boundaries from `BitUnpacker`, and then lazily load each block when it is requested. Only the `get_val` function uses the cache: `get_row_ids_for_value_range` does not (yet), because it would be necessary to partition the row ids by block, and most of the time consumers using it are already loading reasonably large ranges anyway. See paradedb/paradedb#2894 for usage. There are a few 2x speedups in the benchmark suite, as well as a 1.8x speedup on a representative customer query. Unfortunately there are also some 13-19% slowdowns on aggregates: it looks like that is because aggregates use `get_vals`, for which the default implementation is to just call `get_val` in a loop.
## What
Add a `TopDocs::order_by` method, which supports ordering by multiple fast fields and scores in one collection pass, as defined by the `TopOrderable` trait. The `TopOrderable` trait is implemented (by a macro) for tuples of length 1 through 3 (for now).
## How
Add:
* a `TopOrderable` trait which is implemented for tuples, and a `TopOrderableCollector` to collect for it.
* a `Feature` trait which is implemented for `Score`s, and for fast fields.
* To allow for boxing/dynamic dispatch of `Features` (which reduces code generation when the sort columns are not known until runtime), `Arc<dyn Feature>` is implemented via `ErasedFeature`.
* a `TopNCompare` trait which can be used together with a `LazyTopNComputer` to lazily fetch columns during TopN.
* This new interface is necessary because `TopNComputer` does not allow for lazily fetching additional fields for the comparison tuple, which can eliminate a lot of IO when tiebreakers are only rarely actually coming into play in the comparison (because most values are being eliminated by earlier columns).
* It could also allow for making `DocId`/`DocAddress` tiebreaking optional ([see](quickwit-oss#2672 (comment))), via something like a "`DocIdFeature`".
This interface additionally could not use the `CustomScorer` APIs because it does not allow segments to Top-N a different type than their final output type (which is essential for ordering by `String`s).
## Note
This patch isolates everything to one module, but should almost certainly be split up into multiple modules, and better integrated with the existing modules. I was hoping to get some feedback on it before rearranging things, but I'm very happy to do so!
----
Upstream at quickwit-oss#2681
This removes a check against `IndexWriterOptions` which disallowed zero indexing worker threads (`num_worker_threads`).
- Use a bitset to track used buckets in the `SharedArenaHashmap`, allowing for more efficient iteration - Create a global pool for both `MemoryArena` and `IndexingContext` - Reduce the MemoryArea page size by half (it's now 512KB instead of 1MB) - Centralize thread pool instances in `SegmentUpdater` so we can elide making them if all nthread sizes are zero
Adjusts `Feature` to use `Option` where necessary. `StringFeature::SegmentOutput` avoids being wrapped in an `Option` because the `u64::MAX` "niche" is safe to use in that case (since it would require that many terms to trigger a collision). Surprisingly, no performance impact downstream. I additionally needed to adjust the unit tests to: * include a NULL * use a stable id, because for some reason at three segments (but not two?) the fact that segment ords are not stable was exposed.
A MemoryArena should support allocations up to 4GB and #60 broke this by not accounting for the "max page id" when pages are now 50% the size of what the originally were. This cleans up the code so things stay in sync if we change NUM_BITS_PAGE_ADDR again and adds a unit test
rebasedming
marked this pull request as ready for review
September 29, 2025 20:06
mdashti
force-pushed
the
main
branch
2 times, most recently
from
December 6, 2025 06:37
15d7b1c to
a7329cc
Compare
Member
|
Going to close this. We can reopen post-rebase in a brand new PR. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.