[10125] [encode path] Minor optimizations to arrow-flight#10137
Open
Rich-T-kid wants to merge 7 commits into
Open
[10125] [encode path] Minor optimizations to arrow-flight#10137Rich-T-kid wants to merge 7 commits into
Rich-T-kid wants to merge 7 commits into
Conversation
github-actions
Bot
added
arrow
Rich-T-kid
commented
Jun 12, 2026
| } | ||
|
|
||
| /// Place the `FlightData` in the queue to send | ||
| #[inline] |
Contributor
Author
There was a problem hiding this comment.
The compiler very likely could have inlined this, but I think its work adding this explicitly.
Contributor
|
run benchmarks flight |
|
🤖 Arrow criterion benchmark running (GKE) | trigger CPU Details (lscpu)Comparing rich-T-kid/minor-arrow-flight-opt (d02e297) to 826b808 (merge-base) diff File an issue against this benchmark runner |
|
🤖 Arrow criterion benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usagebase (merge-base)
branch
File an issue against this benchmark runner |
Contributor
Author
|
seems like its mostly noise |
Contributor
Author
its interesting that this seems to always regress |
Rich-T-kid
changed the title
Rich-T-kid
force-pushed
the
rich-T-kid/minor-arrow-flight-opt
branch
2 times, most recently
from
2c00600 to
337abd5
Compare
Rich-T-kid
commented
Jun 12, 2026
…mory to fill array data
Rich-T-kid
force-pushed
the
rich-T-kid/minor-arrow-flight-opt
branch
from
337abd5 to
094579b
Compare
Contributor
Author
|
@Jefffrey I meant to ping you on this PR . Sorry about that! |
Contributor
|
run benchmarks flight |
|
🤖 Arrow criterion benchmark running (GKE) | trigger CPU Details (lscpu)Comparing rich-T-kid/minor-arrow-flight-opt (505fb20) to 826b808 (merge-base) diff File an issue against this benchmark runner |
|
🤖 Arrow criterion benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usagebase (merge-base)
branch
File an issue against this benchmark runner |
Contributor
Author
|
Nice, regressions are gone. should re-run when 54faeda gets merged. I expected a larger improvement for larger rows/columns batches. I'll profile & update the PR |
Rich-T-kid
force-pushed
the
rich-T-kid/minor-arrow-flight-opt
branch
from
7c6f70f to
37c7231
Compare
Rich-T-kid
force-pushed
the
rich-T-kid/minor-arrow-flight-opt
branch
from
37c7231 to
166e2e6
Compare
Contributor
Author
|
@alamb could you run the benchmarks for arrow-flight? 🚀 |
Contributor
|
run benchmark flight |
|
🤖 Arrow criterion benchmark running (GKE) | trigger CPU Details (lscpu)Comparing rich-T-kid/minor-arrow-flight-opt (166e2e6) to 826b808 (merge-base) diff File an issue against this benchmark runner |
|
🤖 Arrow criterion benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)Details
Resource Usagebase (merge-base)
branch
File an issue against this benchmark runner |
Contributor
Author
|
similar story here. encode is path shows good improvements but the roundtrip is the same or slightly worse. |
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
5 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.
Which issue does this PR close?
starting small 😄
Rationale for this change
The arrow-flight encode path was allocating intermediate Vecs to hold data that was immediately iterated and discarded. Replacing these with lazy iterators and inlining the one helper that existed only to loop removes allocations that served no purpose beyond bridging two adjacent lines of code.
What changes are included in this PR?
[commit #1]
Impl<Iterator>.[commit #2]
[commit #3]
CompressionContexttoIpcWriteContextand added anfbb: FlatBufferBuilder<'static>field to it[commit #4]
IpcWriteContextgains a scratch:Vec<u8>field. When set before a call toIpcDataGenerator::encode(), the existing allocation is reused instead of allocating a fresh buffer for each batch's arrow data body.FlightIpcEncodermaintains anArrowDataPool, a small pool of Arc<Mutex<Vec<Vec>>> buffers pre-sized to the gRPC message limit (2 MiB). Before eachencode()call, a buffer is acquired from the pool and placed inIpcWriteContext::scratch. After encoding, the buffer is wrapped inPooledBufand handed to Bytes::from_owner; when the Bytes is dropped (after the gRPC frame is sent), the buffer is automatically returned to the pool rather than freed.[commit #5 & 6]
acquiremethod to also pre-allocate 2MB of space in the vectoripc::encode()calls. the buffers are pre-allocated to themax_flight_data_sizeas an estimate. This means no intermediate vector copies to larger vectorscommit [ #7] final commit
arrow-ipc::encode()sink fromIpcBodySink::Write()toIpcBodySink::collect()memcpyto new vectors.memcpy. this is also why the profile shows alot of memcpy or _platform_memmove on mac.output buffer size changes
This PR also changes the size of buffers being output by
split_batch_for_grpc_response()The old algorithm computed n_batches first via ceiling division, then derived rows_per_batch from that:
This evenly distributes rows across chunks, meaning each buffer ends up smaller than max on average. Thus leaving capacity unused.
The new algorithm works directly from the target size:
This packs each buffer as close to max as possible before moving on.
This matters because the output buffers are pre-allocated to max_flight_data_size. Since the allocation cost is already paid upfront, the only cost of filling a buffer is the memcpy itself. As the profiles show, most time is spent serializing RecordBatches to IPC format, doing that for as many rows as possible in one pass, followed by a single large memcpy, is faster than multiple smaller serializations and copies. Leaving pre-allocated capacity unused means splitting work across more messages than necessary, each carrying its own network and serialization overhead.
note: I expect to have to tune this a bit. this is because the size that is used to determine the total size of the record batch isn't exact. strings vary from row to row, so its hard to get the math 100% correct. but the closer we can get to the max_size the better
Are these changes tested?
yes
Are there any user-facing changes?
no