Draft
Conversation
|
Thanks for opening this PR! Here are a few tips to help make the review process smooth for everyone. PR title guidelines
Pre-merge checklist
|
def-
force-pushed
the
pr-pgwire-opt
branch
11 times, most recently
from
be9acf3 to
e175bc9
Compare
…llocation Encode pgwire DataRow messages directly from RowRef datums to BytesMut, eliminating the intermediate Vec<Option<Value>> that values_from_row() allocated per row. For string/bytes columns, this also eliminates the per-value clone (to_owned/to_vec) since we now write directly from the borrowed Datum slice. Adds encode_data_row_direct() which handles the full DataRow wire format (type byte, message length, field count, per-field length + text encoding) in a single pass. A new BackendMessage::PreEncoded variant lets the protocol layer send pre-encoded wire bytes through the codec without re-framing. Benchmark: 1.02-1.24x faster per row, 1.06-1.30x in batch (10k rows). String-heavy rows see the largest improvement (1.30x) due to eliminated string cloning. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…ction Adds encode_datum_binary_direct() that encodes common datum types directly into BytesMut for binary pgwire format, skipping the intermediate Value enum construction. This eliminates heap allocations for string/bytes columns and avoids ToSql trait dispatch for simple types (bool, ints, floats, timestamps, dates, times, intervals, uuids). Benchmark results (5-column rows): - Integers: 137ns → 124ns (1.10x) - Strings: 164ns → 115ns (1.43x) - Mixed: 142ns → 115ns (1.23x) - Timestamps: 185ns → 175ns (1.06x) - Batch strings 10k: 1.74ms → 1.15ms (1.52x) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…ue>> allocation Replace values_from_row() in encode_copy_row_text() and encode_copy_row_csv() with direct datum iteration using encode_datum_text_direct(). This eliminates the per-row Vec<Option<Value>> allocation and avoids cloning string/bytes data. Also adds an escape fast-path for COPY text: when the encoded value contains no special characters (common for numbers, dates, timestamps), copy the entire buffer with extend_from_slice instead of per-byte scanning. Fixes a JSONB encoding bug in encode_datum_text_direct() where the JSONB match arm was positioned after Datum::String/True/False/Numeric arms, causing JSONB values to be formatted as plain scalars instead of JSON. Benchmark: 1.64-2.17x faster COPY text encoding (integers 1.64x, strings 2.17x, mixed 1.89x in batch). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace byte-by-byte scanning in CopyTextFormatParser::consume_raw_value() with memchr::memchr3() which uses SSE2/AVX2 SIMD to find delimiter bytes at 16-32 bytes per CPU cycle instead of 1 byte per 5-6 condition checks. The old code checked is_eof(), is_end_of_copy_marker(), is_end_of_line(), is_column_delimiter(), and peek() for EVERY byte - that's 4-6 bounds checks and comparisons per normal character. The new code finds the next special byte (\t, \n, \\) in a single vectorized pass, then handles it. Benchmark: 3.5x faster for long strings (~80 chars), 1.4x for medium strings (~17 chars), 1.1x for escaped data. Short integer values show no improvement since per-value framework overhead dominates. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
… 2.1-4.6x speedup Two optimizations for the COPY FROM text ingestion path: 1. parse_bool: Replace s.trim().to_lowercase() (heap-allocates a String per call) with length-dispatched ASCII case-insensitive byte matching via eq_ignore_ascii_case. 1.2-1.4x faster per call. 2. COPY FROM text decode: Replace Value::decode_text → into_datum → Vec<Datum> → Row::pack pipeline with direct decode_text_into_row into RowPacker. Eliminates per-row Vec<Datum> allocation, per-field Value construction, and per-text-column String allocation (s.to_owned()). 2.1x faster for int cols, 4.6x faster for text cols in batch. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…alue>> allocation" This reverts commit 4f617c9.
…llocation Encode pgwire DataRow messages directly from RowRef datums to BytesMut, eliminating the intermediate Vec<Option<Value>> that values_from_row() allocated per row. For string/bytes columns, this also eliminates the per-value clone (to_owned/to_vec) since we now write directly from the borrowed Datum slice. Adds encode_data_row_direct() which handles the full DataRow wire format (type byte, message length, field count, per-field length + text encoding) in a single pass. A new BackendMessage::PreEncoded variant lets the protocol layer send pre-encoded wire bytes through the codec without re-framing. Benchmark: 1.1-1.6x faster per row, 1.2-1.4x in batch (10k rows). String-heavy rows see the largest improvement (1.6x) due to eliminated string cloning. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace `Decimal::to_standard_notation_string()` (which heap-allocates both a Vec<u8> for coefficient digits and a String for the result) with `write_numeric_standard_notation()` that writes directly to any fmt::Write with zero heap allocations. The new function extracts digits from the internal `coefficient_units()` representation (a &[u16] slice, no allocation) into a stack-allocated [u8; 39] buffer, builds the complete formatted output in a stack-allocated [u8; 80] buffer, and writes it with a single write_str() call. Updated the two main hot paths: - `format_numeric()` in strconv.rs (pgwire text encoding) - `Datum::Numeric` Display impl in scalar.rs Benchmarks (criterion): - Batch (10k values to shared buffer): 306µs → 118µs (2.6x faster) - Per-value write-only: 1.8x-6.0x faster depending on value - Per-value to new String: 1.0x-2.1x faster Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace `to_standard_notation_string()` (which heap-allocates Vec<u8> + String per value) with `write_numeric_standard_notation()` that writes directly to fmt::Write with zero heap allocations. Uses `coefficient_units()` (&[u16] slice, no alloc) to extract digits into a stack [u8;82] buffer, builds complete output in a stack [u8;200] buffer, and writes with single write_str() call. Benchmark: 2.0-4.3x faster per value, 2.3x faster in batch (10k values). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
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
1 participant
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.
Taken and cleaned up from #35076