Add benchmarks for arrow-avro writer #8165
Conversation
github-actions
bot
added
arrow
- Introduce `avro_writer` benchmark suite in `arrow-avro/benches/avro_writer.rs`. - Test writing performance for various data types, including Boolean, Int32, Int64, Float32, Float64, Binary, TimestampMicrosecond, and Mixed schemas. - Update `Cargo.toml` to include the `avro_writer` benchmark target.
jecsand838
force-pushed
the
avro-writer-benchmarks
branch
from
c1dd2a8 to
936d255
Compare
There was a problem hiding this comment.
Thanks @jecsand838 -- I have some suggestions on how to improve these benchmarks, but I also think we could do that as a follow on
arrow-avro/benches/avro_writer.rs
Outdated
|
|
||
| const SIZES: [usize; 3] = [100, 10_000, 1_000_000]; | ||
|
|
||
| fn make_bool_array(n: usize) -> BooleanArray { |
There was a problem hiding this comment.
Most of our other benchmarks use random input to avoid pathalogical cases due to patterns in the input
I recommend doing the same in this PR
Here are some examples
arrow-rs/arrow/src/util/data_gen.rs
Line 37 in a9b4221
There was a problem hiding this comment.
Ah that makes sense. I can add that in tonight.
There was a problem hiding this comment.
@alamb I just pushed up some changes based on your feedback regarding the random inputs. Let me know what you think!
| .collect() | ||
| }); | ||
|
|
||
| fn ocf_size_for_batch(batch: &RecordBatch) -> usize { |
There was a problem hiding this comment.
It's the size of the Avro Object Container File.
There was a problem hiding this comment.
Also to be a bit more clear, my intention here was to feed this usize byte count value into Throughput::Bytes() so Criterion reports MB/s for actual on‑disk OCF bytes written per iteration.
arrow-avro/benches/avro_writer.rs
Outdated
| let bytes = ocf_size_for_batch(batch); | ||
| group.throughput(Throughput::Bytes(bytes as u64)); | ||
| match rows { | ||
| 10_000 => { |
There was a problem hiding this comment.
A common batch size is 8k as well -- I might suggest testing 4k or 8k batches along with 100k to better represent any per-batch overhead (as with 100K it will be drowned out)
There was a problem hiding this comment.
That's a good call out. I'll push those changes up tonight as well.
There was a problem hiding this comment.
Pushed these changes up as well.
jecsand838
force-pushed
the
avro-writer-benchmarks
branch
from
2f07609 to
c6c6aeb
Compare
|
I took a look at the changes and they look good to me -- thanks @jecsand838 |
Which issue does this PR close?
Rationale for this change
This PR introduces benchmark tests for the
AvroWriterin thearrow-avrocrate. Adding these benchmarks is essential for tracking the performance of the writer, identifying potential regressions, and guiding future optimizations.What changes are included in this PR?
A new benchmark file,
benches/avro_writer.rs, is added to the project. This file contains a suite of benchmarks that measure the performance of writingRecordBatches to the Avro format.The benchmarks cover a variety of Arrow data types:
BooleanInt32andInt64Float32andFloat64BinaryTimestamp(Microsecond precision)These benchmarks are run with varying numbers of rows (100, 10,000, and 1,000,000) to assess performance across different data scales.
Are these changes tested?
Yes, this pull request consists entirely of new benchmark tests. Therefore, no separate tests are needed.
Are there any user-facing changes?
NA