151 merged changes. Check any of them.
Since June, the maintainers of 36 projects have merged 151 changes that started with Perfloop. Go, PostgreSQL®, Redis® and Elasticsearch are among them.
36 projects132 pull requests, 6 commits and 13 project fixes of issues Perfloop reported
Cumulative merged changes from Jun 21 to Nov 18, 2026, reaching 151. The list at the end of the page holds every change.
Each square is one merged change, in the colour of its pattern. Point at one to read it; select it to open the merge on GitHub.
Merged per month
Through September, 51 of the 53 merged changes were performance gains. Since October 1, 71 bug fixes have merged.
Agents write the software. Perfloop makes it self-improving. It finds what is slow, wrong or unsafe and proves every change.
It learns from the answers it gets. Perfloop reads maintainer review comments, closed pull requests and failed checks, and keeps each lesson with the code it came from.
Ten projects you know
Ten of the 36, each with one of its merged changes.
Go
Programming language, 139,957 stars on GitHubA recover in a range-over-func loop body returned nil after the goroutine stack grew or shrank. Merged Nov 10, by Cuong Manh Le (their fix, Perfloop's report)
bug fixedPostgreSQL®
Relational database, 22,843 stars on GitHubFreeing a WAL reader leaked separately allocated records still in its decode queue. Merged Nov 4, by Fujii Masao
bug fixedRedis®
In-memory data store, 77,520 stars on GitHubAfter the IDMP expiry cron pruned a stream, CLUSTER SLOT-STATS kept counting the freed memory. Merged Oct 28, by Sergei Georgiev
bug fixedElasticsearch
Search engine, 78,689 stars on GitHubAsync cross-cluster searches returned the wrong page of hits while a remote cluster was still running. Merged Nov 11, by Eugene Rizhkov
bug fixedPrometheus®
Monitoring system, 67,647 stars on GitHubPromQL® range functions built a fresh one-sample vector for every scalar argument, series and query step. 6 changes merged. This one: Nov 2, by Bartlomiej Plotka
2.1× lowercurl
Data transfer tool and library, 43,527 stars on GitHubSince curl 8.18.0, cookie-file imports could overwrite cookies added with CURLOPT_COOKIELIST before the first transfer, contrary to their documented priority. 2 changes merged. This one: Nov 14, by Daniel Stenberg
bug fixedCoreDNS
The default DNS server in Kubernetes®, 14,433 stars on GitHubThe cache cut fractional SERVFAIL lifetimes to whole seconds: a 500ms failure expired at once, and a 1.5s failure could be retried after one second. 4 changes merged. This one: Nov 15, by Yong Tang
bug fixedLiveKit
Real-time audio and video server, 22,235 stars on GitHubForwarding packets with common 4-byte dependency descriptors allocated memory for their headers. 2 changes merged. This one: Sep 30, by cnderrauber
allocations to zeroDistribution
Container image registry, 10,690 stars on GitHubA cold read of a small blob reserved a 4 MiB buffer. 6 changes merged. This one: Nov 10, by Milos Gajdos
754× lowerDolt
SQL database with version control, 24,908 stars on GitHubUndrained keyless secondary-index scans left producer goroutines blocked on full channels. 2 changes merged. This one: Oct 13, by Nick Tobey
bug fixedWhat it finds
127 of the 151 changes are in Go. The rest are in Rust, C, Python, C++, C# and Java.
How large the measured gains are, one square for each change
The largest measured gains in open source projects
Before against after, on the measurement the change itself reports. Bar length is on a log scale. Most are component benchmarks, and the line under each name says what was measured.
One change, from finding to merge
parseablehq/parseable#1846, merged Nov 9 by Nikhil Sinha.
found
Each streamed SQL batch was written to JSON by Arrow, parsed back into a value tree and serialized a second time.
changed
Perfloop returns Arrow's JSON buffer for each streamed batch directly, keeping the newline framing and field-name prefix.
measured
Streamed SELECT * over 1,000,000 rows (346 MB, 53 batches) · 7 interleaved pairs
On your code
The same process that produced this record, on the repositories you choose.
- 1
You connect the repositories you choose.
Perfloop works only on the repositories an admin selects. - 2
It looks for what is slow, wrong or unsafe.
Performance problems, bugs and security flaws. - 3
It proves each change before you see it.
By a measurement, or by a test that asserts the fix. - 4
It opens a pull request.
The evidence is in its description. sign up →
Tigris Data runs Perfloop on its storage systems
Tigris Data met Perfloop as a pull request in its open source gateway, tigrisdata/tag. Ovais Tariq merged it and asked for access.
Tigris is now a customer. It has merged 48 of Perfloop's pull requests in its public repositories, 34 in tag and 14 in ocache, and wrote 4 fixes of its own for problems Perfloop reported.
merged changes in its public repositories, one square each
The people who merged them
36 maintainers pressed merge. The number is how many each one merged; a fix they wrote themselves for a problem Perfloop reported is named as their own.
- Cuong Manh LeGo own fix
- Eugene RizhkovElasticsearch 1
- Sergei GeorgievRedis® 1
- Arve KnudsenPrometheus® 3
- Bartlomiej PlotkaPrometheus® 3
- Micha ReiserRuff 1
- Alexey MilovidovClickHouse 1
- Daniel Stenbergcurl 2
- Nick TobeyDolt 1
- Aaron SonDolt 1
- Fujii MasaoPostgreSQL® 1
- cnderrauberLiveKit 1
- Raja SubramanianLiveKit 1
- Ruben FiszelWindmill own fix
- Aleix Conchillo FlaquéPipecat 2 + own fix
- Asdine El HrychyWeaviate 1
- Yong TangCoreDNS 4
- Yann ColletLZ4 1
- Milos GajdosDistribution 6
- Florian RapplAngleSharp 1
- Laurent DemaillyFortio, grol 4
- Mustafa AbdelrahmanSkipper 2
- Sandor SzücsSkipper 1
- Daniel LemireRoaringBitmap 8 + own fix
- Guy A MolinariRoaringBitmap 1
- Nikhil SinhaParseable 1
- Keegan Carruthers-SmithZoekt 7
- Seif LotfyAxiom FSST, Axiom HyperLogLog 3
- Liz Fong-Jonespierrec/lz4 1
- Achilleparquet-go, kway-go 13
- Josh Stevensrindexer 1
- Torbjörn Einarssonmp4ff 5 + 4 own fixes
- Jitendra Adhikarigronx 1
- Nuno Lopescpp2rust 8
- AnoopDataFusion DuckLake 1
- Ovais TariqTigris TAG, Tigris ocache 48 + 4 own fixes
Read each merged change
Newest first, with the person who merged it.
| Merged | Change | What was wrong | Result | Merged by |
|---|---|---|---|---|
| Nov 18 | Weaviate #13940 | A recoverable queue write failure could leave an HFresh posting marked pending with no task, so later searches did not schedule its merge. | bug fixed | Asdine El Hrychy |
| Nov 18 | CoreDNS #8774 | CoreDNS exited at startup when a valid zone file had an SOA serial of zero, and the auto plugin skipped such a zone. | bug fixed | Yong Tang |
| Nov 18 | CoreDNS #8776 | A secondary whose zone had expired still answered zone transfers with the obsolete records. | bug fixed | Yong Tang |
| Nov 18 | CoreDNS #8778 | With a hostname upstream, forward could refuse to start or configure an incomplete upstream list when the resolver truncated its UDP answer. | bug fixed | Yong Tang |
| Nov 18 | parquet-go #685 | After two row seeks in a row, a file reader could return the wrong row or skip later pages. | bug fixed | Achille |
| Nov 17 | parquet-go #683 | A descending sort of an optional column put the nulls on the wrong side, so the rows disagreed with the sorting columns the file declares. | bug fixed | Achille |
| Nov 17 | mp4ff #721 | Streamed sample getters read every sample from the file on its own, even in a contiguous run. | 2.5× lower | Torbjörn Einarsson |
| Nov 17 | LiveKit #5213 | A receiver restart during a retransmission could pair old outgoing packet headers with new media. | bug fixed | Raja Subramanian |
| Nov 17 | mp4ff #720 | CBCS encryption succeeded with an IV different from the init segment's constant IV, and readers then recovered different media bytes. | bug fixed | Torbjörn Einarsson |
| Nov 16 | curl #23945 | A failed allocation while replacing a cookie could discard the old cookie and leave the cookie count one too high. | bug fixed | Daniel Stenberg |
| Nov 16 | cpp2rust #583 | Translated try_emplace calls skipped the side effects of their arguments when the key already existed. | bug fixed | Nuno Lopes |
| Nov 16 | Zoekt #1250 | A failed blob read during git indexing could publish a partial or empty shard over the complete index, so files dropped out of search. | bug fixed | Keegan Carruthers-Smith |
| Nov 15 | parquet-go #640 | Schema.Reconstruct could fill a missing column with a previous call's values, or let a stale value satisfy a required field. | bug fixed | Achille |
| Nov 15 | parquet-go #632 | Packing many small Parquet files allocated a 1,024-value scratch buffer for every segment and column. | 1.6× lower | Achille |
| Nov 15 | parquet-go #638 | variant.Marshal took the pooled encoder path even for a single integer, a short string or an empty container. | 2.4× lower | Achille |
| Nov 15 | cpp2rust #616 | A translated C++ string left no terminator after a move, so size() wrapped and an unsafe substring length could underflow. | bug fixed | Nuno Lopes |
| Nov 15 | CoreDNS #8756 | The cache cut fractional SERVFAIL lifetimes to whole seconds: a 500ms failure expired at once, and a 1.5s failure could be retried after one second. | bug fixed | Yong Tang |
| Nov 14 | curl #23942 | Since curl 8.18.0, cookie-file imports could overwrite cookies added with CURLOPT_COOKIELIST before the first transfer, contrary to their documented priority. | bug fixed | Daniel Stenberg |
| Nov 14 | rindexer #474 | Address-filtered log requests that spanned several address chunks could rewind the historical cursor and dispatch the same logs again. | bug fixed | Josh Stevens |
| Nov 14 | mp4ff #706 | HEVC slice headers that reused the SPS's only short-term reference picture set or long-term picture returned the wrong reference pictures. | bug fixed | Torbjörn Einarsson |
| Nov 13 | mp4ff a72c55d | Boxes decoded while streaming, such as emsg message data and avcC parameter sets, were overwritten by the next box read.Reported by Perfloop in Eyevinn/mp4ff#712. | bug fixed | Torbjörn Einarsson (their fix, Perfloop's report) |
| Nov 13 | mp4ff b308adc | Streamed sample access gave a track the first track's trex defaults, such as the video duration 3000 instead of 1024 for audio samples.Reported by Perfloop in Eyevinn/mp4ff#711. | bug fixed | Torbjörn Einarsson (their fix, Perfloop's report) |
| Nov 13 | mp4ff a95a8f5 | SttsBox.GetTimeCode wrapped the elapsed time of a sample that started more than 2^32 ticks into a track.Reported by Perfloop in Eyevinn/mp4ff#708. | bug fixed | Torbjörn Einarsson (their fix, Perfloop's report) |
| Nov 13 | Dolt #12427 | GC admission could drop a cancellation signal, because the broadcast could land before the listener registered. | bug fixed | Aaron Son |
| Nov 13 | Zoekt #1230 | Case-insensitive substring verification could lose matches that differ only within a Unicode fold class, such as sss and ssſ or σσσ and σσς. | bug fixed | Keegan Carruthers-Smith |
| Nov 12 | Zoekt #1227 | A failed shard rename during index publication could delete the prior index data or hide the error. | bug fixed | Keegan Carruthers-Smith |
| Nov 12 | RoaringBitmap #591 | On arm64, batch iteration over a bitmap container found each set bit with a population count, which has no scalar instruction there.Reported by Perfloop in RoaringBitmap/roaring#553. | gain, not sized | Daniel Lemire (fix by Liz Fong-Jones, Perfloop's report) |
| Nov 12 | Distribution #5069 | A blob upload could answer 202 Accepted with an offset the stored file did not reach, so the next request got 416. | bug fixed | Milos Gajdos |
| Nov 12 | gronx #81 | A task admitted just before Stop or the cutoff left its result goroutine waiting forever. | bug fixed | Jitendra Adhikari |
| Nov 12 | mp4ff #705 | HEVC slice headers whose reference picture set was predicted from another set failed to parse or returned the wrong reference pictures. | bug fixed | Torbjörn Einarsson |
| Nov 11 | pierrec/lz4 #297 | lz4c left an old file's tail after a shorter output and could write its output over its own input. | bug fixed | Liz Fong-Jones |
| Nov 11 | Distribution #4906 | Manifest uploads checked each referenced layer in storage one at a time. | 6.0× lower | Milos Gajdos |
| Nov 11 | Distribution #5072 | A pull-through cache sent the full blob again to clients that already held it, because its first ETag was unquoted. | gain, not sized | Milos Gajdos |
| Nov 11 | Distribution #5071 | Each resumed upload chunk saved the same digest checkpoint to storage twice. | gain, not sized | Milos Gajdos |
| Nov 11 | Elasticsearch #163290 | Async cross-cluster searches returned the wrong page of hits while a remote cluster was still running. | bug fixed | Eugene Rizhkov |
| Nov 10 | Go f75d503 | A recover in a range-over-func loop body returned nil after the goroutine stack grew or shrank.Reported by Perfloop in golang/go#82930. | bug fixed | Cuong Manh Le (their fix, Perfloop's report) |
| Nov 10 | Distribution #5070 | A cold read of a small blob reserved a 4 MiB buffer. | 754× lower | Milos Gajdos |
| Nov 9 | Parseable #1846 | Streamed queries decoded every batch's JSON into a value tree and encoded it again before sending it. | 4.5× lower | Nikhil Sinha |
| Nov 8 | cpp2rust #614 | Pointers to fields of temporary objects outlived their owners in refcount translations, causing dangling-pointer panics. | bug fixed | Nuno Lopes |
| Nov 7 | LZ4 #2066 | LZ4F_decompress reported zero bytes consumed and produced when it failed, even after it had written output. | bug fixed | Yann Collet |
| Nov 4 | PostgreSQL® 6a146d2 | Freeing a WAL reader leaked separately allocated records still in its decode queue. | bug fixed | Fujii Masao |
| Nov 3 | cpp2rust #584 | Translated getsockname on Unix sockets returned success with the caller's address buffer unchanged. | bug fixed | Nuno Lopes |
| Nov 3 | mp4ff #690 | aac.NewADTSHeader accepted a payload of 8185 bytes and encoded a frame length of zero.Reported by Perfloop in Eyevinn/mp4ff#689. | bug fixed | Torbjörn Einarsson (their fix, Perfloop's report) |
| Nov 3 | mp4ff #686 | HEVC parameter-set extraction dropped the final set when an Annex-B stream ended before video, and appending to one returned set could overwrite another. | bug fixed | Torbjörn Einarsson |
| Nov 3 | Skipper #4396 | Warm HTTP cache reads copied each serialized response before decoding it. | −20% | Mustafa Abdelrahman |
| Nov 3 | Prometheus® #20180 | PromQL® quantiles multiplied equal samples by interpolation weights, so the smallest subnormal value came back as zero. | bug fixed | Bartlomiej Plotka |
| Nov 3 | Prometheus® #20184 | Streamed Remote Read returned in-order head samples that its isolation snapshot should have hidden. | bug fixed | Bartlomiej Plotka |
| Nov 2 | cpp2rust #581 | Unlabelled breaks inside nested goto and switch macros left a generated loop instead of their intended target. | bug fixed | Nuno Lopes |
| Nov 2 | cpp2rust #580 | Translated repeated-character string append panicked on a checked borrow for strings whose data pointer was used. | bug fixed | Nuno Lopes |
| Nov 2 | Prometheus® #19389 | PromQL® range functions built a fresh one-sample vector for every scalar argument, series and query step. | 2.1× lower | Bartlomiej Plotka |
| Nov 1 | cpp2rust #577 | Translated substring calls evaluated the receiver and arguments more than once, changing the selected bytes and caller state. | bug fixed | Nuno Lopes |
| Oct 29 | Zoekt #1202 | Chunk search reuses resolved lines when building ranges and scoring matches. | −20% | Keegan Carruthers-Smith |
| Oct 28 | Windmill #12087 | Every code job deep-copied the cached script's module map: path keys, module contents and lock strings.Reported by Perfloop in windmill-labs/windmill#12084. | 1.8× lower | Ruben Fiszel (their fix, Perfloop's report) |
| Oct 28 | Redis® #15887 | After the IDMP expiry cron pruned a stream, CLUSTER SLOT-STATS kept counting the freed memory. | bug fixed | Sergei Georgiev |
| Oct 24 | Axiom HyperLogLog #75 | Every dense estimate called a general math.Pow on every register. | 5.7× lower | Seif Lotfy |
| Oct 23 | AngleSharp #1388 | HTML serialization allocated pooled builders to rebuild strings even when no characters needed escaping. | −23% | Florian Rappl |
| Oct 22 | E2B #3713 | Killing a sandbox that its node had already dropped subtracted the sandbox's resources from the node's allocations a second time. | bug fixed | the project |
| Oct 22 | Tigris TAG #274 | Cached multipart-completion replays were not bound to the access key that completed the upload.Reported by Perfloop in tigrisdata/tag#242. | bug fixed | Ovais Tariq (their fix, Perfloop's report) |
| Oct 22 | Tigris TAG #269 | A truncated whole-object cache entry was served with the full Content-Length until its TTL ran out.Reported by Perfloop in tigrisdata/tag#249. | bug fixed | Ovais Tariq (their fix, Perfloop's report) |
| Oct 22 | Tigris TAG #271 | In signing mode, TAG forwarded signed streaming uploads without a check of each chunk's signature.Reported by Perfloop in tigrisdata/tag#243. | bug fixed | Ovais Tariq (their fix, Perfloop's report) |
| Oct 21 | Tigris TAG #248 | Transparent mode treated successful writes and non-read responses as permission to read private cached data. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #256 | Embedded storage closed while detached cache writers were still running during shutdown. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #219 | Broadcasting copied every upstream body chunk even when serving a single listener. | 2.2× lower | Ovais Tariq |
| Oct 21 | Tigris TAG #238 | Bulk delete responses could finish before cache invalidation cleared stale racing refills. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #223 | Warm block-cached multi-block range requests ran serial block presence probes before reading payload data. | 13× lower | Ovais Tariq |
| Oct 21 | Tigris TAG #266 | Transparent local auth rejected URL-only SigV4 presigned GET requests lacking an X-Amz-Content-Sha256 header. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #270 | A stalled cache read from a remote peer held a request until the client disconnected or the server's five-minute write timeout fired.Reported by Perfloop in tigrisdata/tag#222. | bug fixed | Ovais Tariq (their fix, Perfloop's report) |
| Oct 21 | Tigris TAG #251 | Positive short cache range reads were accepted as complete hits before reaching advertised Content-Length. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #247 | Unreadable or malformed legacy tombstones allowed stale metadata and ETags to overwrite newer completions. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #234 | Sparse range requests probed every covering block in the cache even after exceeding fan-out limits. | 3.3× lower | Ovais Tariq |
| Oct 21 | Tigris TAG #258 | The embedded cache closed during shutdown before admitted HTTP handlers finished synchronous cleanup. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #263 | Broadcaster terminal error loss published incomplete response bodies under complete-object cache metadata. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #245 | Incomplete upstream streamed responses finished framing around received prefixes instead of aborting. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #206 | Copying response headers lowercased every ordinary header key and allocated memory for each one. | −5.5% | Ovais Tariq |
| Oct 21 | Tigris TAG #259 | Creating the HTTP server inside Start allowed early Stop calls to miss the uninitialized server. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #265 | Every multi-block Range hit ran a serial presence preflight probe before serving cached data. | 2.2× lower | Ovais Tariq |
| Oct 21 | Tigris TAG #239 | Parquet footer warming stopped when a metadata-only HEAD entry was already visible after multipart completion. | to zero | Ovais Tariq |
| Oct 21 | Tigris TAG #237 | Read-triggered Parquet footer scans lacked concurrency bounds, creating unrestricted background workers during burst tail reads. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #240 | Block fetches continued admitting tasks and creating state after parent context cancellation. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #267 | Signing mode rejected valid SDK SigV4 requests when metadata headers contained repeated ASCII spaces. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #257 | Tiered PUTs routed to the local engine acknowledged and served bodies with mismatched payload hashes. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #215 | Single-range cache hits relayed chunks through an intermediate pipe before writing to the response. | 2.2× lower | Ovais Tariq |
| Oct 21 | Tigris TAG #264 | Failed whole-body cache writes left relay goroutines and populate reservations alive. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #253 | Delayed tiered background re-tier tasks could overwrite newer metadata from local writes. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #260 | Full-object GET requests hung when upstream transport failed before sending response headers. | bug fixed | Ovais Tariq |
| Oct 21 | Tigris TAG #261 | Late GET requests that missed the subscription window could mix upstream response headers and bodies. | bug fixed | Ovais Tariq |
| Oct 19 | grol #544 | Negative range bounds that underflowed caused panics instead of returning bounded slices. | bug fixed | Laurent Demailly |
| Oct 19 | grol #545 | Failed compound assignments stored arithmetic errors into identifiers, array elements, or map fields. | bug fixed | Laurent Demailly |
| Oct 19 | cpp2rust #544 | Translated map subscripting used equality instead of ordering equivalence, returning wrong values or replacing entries. | bug fixed | Nuno Lopes |
| Oct 14 | Distribution #4921 | Canonical full-hash fallback validation executed a duplicate SHA-256 pass over stored blob bytes. | 1.9× lower | Milos Gajdos |
| Oct 13 | Dolt #12066 | Undrained keyless secondary-index scans left producer goroutines blocked on full channels. | bug fixed | Nick Tobey |
| Oct 12 | Tigris TAG #252 | CopyObject requests with no-store directives still triggered post-copy cache warming and origin reads. | 15× lower | Ovais Tariq |
| Oct 11 | Fortio #1167 | FastClient reused connections after truncating oversized responses, parsing remaining body bytes as subsequent responses. | bug fixed | Laurent Demailly |
| Oct 11 | Fortio #1168 | FastClient spun in empty-buffer read loops when chunks exceeded the response buffer. | bug fixed | Laurent Demailly |
| Oct 9 | DataFusion DuckLake #368 | Every row-ID scan filtered out deletes that did not exist. | 2.5× lower | Anoop |
| Oct 8 | Skipper #4326 | Default text access logs built and copied field maps even without supplemental fields. | −15% | Sandor Szücs |
| Oct 6 | E2B #3671 | A sandbox create kept waiting for a fresh network slot while a returned one sat ready. | bug fixed | the project |
| Oct 6 | Skipper #4324 | The steady-state fade-in path copied every endpoint on each load-balanced request. | 1.5× lower | Mustafa Abdelrahman |
| Sep 30 | LiveKit #4681 | Forwarding packets with common 4-byte dependency descriptors allocated memory for their headers. | to zero | cnderrauber |
| Sep 25 | Pipecat #5906 | RTVIObserver kept the ID of every frame it saw, audio frames included, for the life of a session.Reported by Perfloop in pipecat-ai/pipecat#5882. | bug fixed | Aleix Conchillo Flaqué (their fix, Perfloop's report) |
| Sep 21 | Prometheus® #19717 | OTLP explicit histogram conversion repeatedly reconstructed metric suffix strings inside data-point and bucket loops. | −18% | Arve Knudsen |
| Sep 17 | Pipecat #5811 | AudioVolumeTracker rebuilt the entire rolling window buffer on every audio chunk. | 1.9× lower | Aleix Conchillo Flaqué |
| Sep 16 | Pipecat #5805 | VADController broadcast redundant UserSpeakingFrame events downstream on every audio frame while speaking. | −16% | Aleix Conchillo Flaqué |
| Sep 16 | Axiom FSST #7 | Table deserialization rebuilt the default lookup state right before overwriting it from the stream. | 3.0× lower | Seif Lotfy |
| Sep 9 | Tigris ocache #253 | YCSB key preloading ran serially, ignoring the configured concurrency during the load phase. | 5.1× lower | Ovais Tariq |
| Sep 9 | Tigris ocache #252 | Sub-second ring heartbeats triggered redundant CAS retries because ring timestamps use one-second precision. | 2.2× lower | Ovais Tariq |
| Sep 9 | Tigris ocache #245 | Segment writes allocated an extra copy buffer for SectionReader inputs during recompaction. | −31% | Ovais Tariq |
| Sep 9 | Tigris ocache #220 | File lock lookups allocated a throwaway mutex on every lookup for an existing path. | −10% | Ovais Tariq |
| Sep 9 | Tigris ocache #273 | YCSB latency reporting copied every per-operation latency slice before sorting percentile values. | 1,015× lower | Ovais Tariq |
| Sep 3 | Tigris TAG #198 | Background full-object cache warms routed through broadcaster listeners instead of streaming directly to cache writers. | 1.6× lower | Ovais Tariq |
| Sep 3 | Tigris TAG #197 | Large cache hits staged data through an io.Pipe goroutine and extra copy. | 1.8× lower | Ovais Tariq |
| Sep 3 | Tigris TAG #199 | Transparent forwarding deep-cloned every inbound header map and value slice on each uncached S3 operation. | −15% | Ovais Tariq |
| Aug 28 | Tigris ocache #241 | Local streamed Puts crossing the inline threshold bypassed FileManager's pooled copy buffer. | 8.9× lower | Ovais Tariq |
| Aug 28 | Tigris ocache #244 | Inline key-value listings allocated and copied protobuf-decoded values a second time. | 2.0× lower | Ovais Tariq |
| Aug 27 | RoaringBitmap #557 | Bitmap.ToArray decoded dense bitmap containers one set bit at a time.Reported by Perfloop in RoaringBitmap/roaring#556. | 1.8× lower | Daniel Lemire (their fix, Perfloop's report) |
| Aug 24 | RoaringBitmap #555 | Non-inplace array subtraction converted array containers into bitmap containers and scanned them. | 4.2× lower | Daniel Lemire |
| Aug 24 | Prometheus® #19503 | Sampled remote reads cloned and merged response labels even when no external labels were configured. | −16% | Arve Knudsen |
| Aug 23 | Tigris ocache #240 | Delete and overwrite cleanup decoded unused inline payload data when reading value messages. | 1.7× lower | Ovais Tariq |
| Aug 22 | Tigris ocache #237 | Every topology response allocated and sorted a copy of normalized token slices. | 1.9× lower | Ovais Tariq |
| Aug 22 | Tigris ocache #238 | Cleaner scans decoded full inline payloads and control messages for ordinary non-expiring rows. | 110× lower | Ovais Tariq |
| Aug 20 | Tigris TAG #175 | Normal PUT forwarding cloned upstream response headers even when it did not retain them. | to zero | Ovais Tariq |
| Aug 20 | Tigris ocache #234 | Discarded YCSB reads materialized unused result slices in memory. | 1.6× lower | Ovais Tariq |
| Aug 20 | Tigris TAG #173 | Cache-rejected PUT requests still ran the full write-through tee admission path and upstream metadata HEAD. | 25× lower | Ovais Tariq |
| Aug 20 | Tigris TAG #172 | Canonical query construction formatted an intermediate string for each escaped byte. | −17% | Ovais Tariq |
| Aug 19 | RoaringBitmap #551 | Two-pointer array intersection loops executed repeated input bounds checks on every element read. | −21% | Daniel Lemire |
| Aug 18 | Tigris ocache #219 | File compaction and segment recompaction copied payloads with no shared throughput bound. | bug fixed | Ovais Tariq |
| Aug 18 | Tigris TAG #164 | Signing endpoint URLs were serialized and reparsed on every request. | −9% | Ovais Tariq |
| Aug 17 | Tigris ocache #217 | Remote GetRangeStream copied every gRPC response chunk into an intermediate buffer before delivery. | −19% | Ovais Tariq |
| Aug 15 | Tigris ocache #210 | Gated access-time reads used Peek, evicting hot keys from the gate cache under capacity pressure. | 1.5× lower | Ovais Tariq |
| Aug 14 | Zoekt #1127 | Code search walked every branch bitmap to reject shards it could rule out up front. | 24× lower | Keegan Carruthers-Smith |
| Aug 14 | kway-go #7 | Two-input MergeSliceFunc discarded comparator signs from failed lookaheads and gallop boundaries. | 1.5× lower | Achille |
| Aug 13 | kway-go #6 | Buffers allocated backing arrays that empty sources and sources ending at growth boundaries never used. | −29% | Achille |
| Aug 13 | ClickHouse #113899 | ClickHouse checked whether constant columns were sorted. Every insert, every row. | gain, not sized | Alexey Milovidov |
| Aug 12 | kway-go #5 | Zero-copy passthrough batches repeatedly grew the merge aggregation buffer and allocated unused capacity. | 2.2× lower | Achille |
| Aug 12 | Prometheus® #19388 | The OTLP write handler built fresh scratch state for every request. | −27% | Arve Knudsen |
| Aug 11 | RoaringBitmap #542 | Lazy bitmap unions cloned and overwrote the receiver container while repeating slice bounds checks. | −15% | Daniel Lemire |
| Aug 11 | RoaringBitmap #539 | Parallel BSI query paths called run-optimization on empty worker-local result bitmaps across workers. | 14× lower | Daniel Lemire |
| Aug 11 | Tigris TAG #148 | Cacheable full-GET requests blocked handler completion until detached cache population finished. | 177× lower | Ovais Tariq |
| Aug 7 | Zoekt #1110 | Unlimited sharded searches repeatedly sorted and truncated matches across every intermediate result chunk. | 26× lower | Keegan Carruthers-Smith |
| Jul 27 | Axiom FSST #5 | Candidate selection during training maintained a min-heap and a separate output list. | −5.7% | Seif Lotfy |
| Jul 24 | parquet-go #581 | Reconstruction of wide maps with repeated descendants allocated 347 times per operation. | −19% | Achille |
| Jul 24 | Ruff #27159 | Cached auto-import completions spent extra time in scope-member lookups. | −14% | Micha Reiser |
| Jul 23 | parquet-go #579 | Default-page INT64 bounds calculations lacked AVX-512VL acceleration. | −27% | Achille |
| Jul 22 | RoaringBitmap #535 | Interleaved source keys during bitmap union shifted metadata slices once for every inserted key. | 6.9× lower | Daniel Lemire |
| Jul 13 | RoaringBitmap #534 | Dense bitmap container materialization decoded set bits using popcount and isolate-lowest-bit operations. | −30% | Daniel Lemire |
| Jul 10 | RoaringBitmap #533 | Scattered IN-list queries in BSI BatchEqual materialized a large number of intermediate bitmaps. | 1.8× lower | Daniel Lemire |
| Jul 9 | Zoekt #1083 | Case-insensitive match verification decoded UTF-8 and performed Unicode lowercasing lookups for every byte. | 2.3× lower | Keegan Carruthers-Smith |
| Jul 7 | parquet-go #550 | variant.Marshal built an intermediate variant.Value tree before encoding it into bytes. | 4.2× lower | Achille |
| Jul 2 | RoaringBitmap #532 | BSI BatchEqual scanned columns with individual binary searches across roaring bitmap bitplanes. | 31× lower | Guy A Molinari |
| Jun 25 | parquet-go #547 | Single-column row comparisons incurred interface dispatch and copied two 24-byte Value structures per comparison. | 1.7× lower | Achille |
| Jun 21 | parquet-go #542 | Writing time columns boxed each value individually, allocating on the heap for every row. | 3.8× lower | Achille |
Merge dates and the people who merged were read from GitHub on November 18, 2026.