Batch no-run container reads during decoding

perfloop/roaring · Chatty I/O

https://perfloop.ai/t/oss/case_bc0tc0yh01

Verdict

VERIFIED · settled 2026-11-14

What happened: The paired measurements met the required improvement.

Hypothesis

A counting-reader probe serialized 16,384 singleton high-key containers and decoded them with Bitmap.ReadFrom through an ordinary io.Reader. It completed the round trip and recorded 16,388 direct Read calls for 163,848 bytes. This check establishes call multiplicity only; it does not measure syscall or storage time.

Bitmap.ReadFrom wraps a reader that is not internal.ByteInput in ByteInputAdapter. ByteInputAdapter.Next allocates the requested slice and invokes io.ReadAtLeast. readFrom calls Next once for each container after the descriptive and offset headers. Thus a no-run bitmap with N low-cardinality containers generates N small reader operations per decode in addition to headers; the parser admits up to 65,536 containers and a singleton array payload is two bytes. The cost is only plausible when the supplied reader has no lower coalescing layer and Read reaches a file, socket, or other storage boundary; ByteInput callers and already-buffered readers do not establish the same cost.

For the no-run format, keycard determines every array or bitmap payload length. Read exact, bounded groups of whole container payloads after the offset table and install sub-slices in each container, so the parser does not overread a caller's stream. Reproduce the counting-reader test with unbuffered file and short-read readers, then benchmark sparse container counts; the expected confirmation is reader or syscall calls falling from N plus a constant per decoded bitmap to one per bounded group while output and post-bitmap reader position remain unchanged.

Change to test: Special-case no-run data in roaringArray.readFrom: calculate payload lengths from keycard and issue exact bounded Next calls for contiguous whole-container groups. Slice each returned group into its containers and keep every group boundary at a serialized container boundary to preserve stream position and limit retained backing storage.

Where it lives

perfloop/roaring · roaring.go

Evidence

Sparse file decode, 1024 containers · 15 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 2380732 68213 −96.9% (−2307313) −2729158 to −2269449 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
read-calls/op 1028 5 −99.5% (−1023) −1023 to −1023 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
B/op 56411 56412 +1 −2 to +4 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 2054 1031 −49.8% (−1023) −1023 to −1023 loss ≤ 0.05 × |baseline median| PASSED

Sparse short decode, 1024 containers · 15 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 2451285 161343 −93.3% (−2287721) −2578422 to −2202116 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
read-calls/op 1058 42 −96% (−1016) −1016 to −1016 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
B/op 56411 56410 −1 −3 to +8 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 2054 1031 −49.8% (−1023) −1023 to −1023 loss ≤ 0.05 × |baseline median| PASSED

Sparse file decode, 4096 containers · 15 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 9490869 196067 −98% (−9298623) −9711485 to −8969137 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
read-calls/op 4100 5 −99.9% (−4095) −4095 to −4095 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
B/op 217214 217215 −1 −12 to +3 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 8198 4103 −50% (−4095) −4095 to −4095 loss ≤ 0.05 × |baseline median| PASSED

Sparse short decode, 4096 containers · 16 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 9757099 564351 −94.2% (−9191077) −10648485 to −8953624 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
read-calls/op 4226 162 −96.2% (−4064) −4064 to −4064 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
B/op 217214 217220 −2.5 −11 to +19 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 8198 4103 −50% (−4095) −4095 to −4095 loss ≤ 0.05 × |baseline median| PASSED

Sparse file decode, 16384 containers · 16 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 38033530 722036 −98.1% (−37322140) −58586528 to −35275675 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
read-calls/op 16388 5 −100% (−16383) −16383 to −16383 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
B/op 868618 868645 +24 −56 to +42 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 32775 16392 −50% (−16383) −16383 to −16383 loss ≤ 0.05 × |baseline median| PASSED

Sparse short decode, 16384 containers · 16 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 38761396 2066487 −94.6% (−36658963) −38590719 to −35326061 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
read-calls/op 16898 642 −96.2% (−16256) −16256 to −16256 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
B/op 868624 868582 −40 −50 to +71 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 32775 16392 −50% (−16383) −16384 to −16383 loss ≤ 0.05 × |baseline median| PASSED

Run-format decode, 6500000 values · 31 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 10136 10119 −23 −378 to +484 loss ≤ 0.05 × |baseline median| PASSED
B/op 14006 14006 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 207 207 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

Weather dataset UnmarshalBinary · 16 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 2490793 1996033 −20.2% (−501906) −691606 to −282120 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
B/op 9150283 9130448 −0.2% (−19836) −19856 to −19812 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 6916 4360 −37% (−2556) −2556 to −2556 loss ≤ 0.05 × |baseline median| PASSED

Weather dataset bufio · 17 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 6644608 3260322 −51.1% (−3395067) −3843454 to −2925255 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
B/op 9140666 9120855 −0.2% (−19839) −19921 to −19787 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 6716 4160 −38.1% (−2556) −2556 to −2556 loss ≤ 0.05 × |baseline median| PASSED

US census dataset UnmarshalBinary · 17 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 189852 147337 −22.2% (−42088) −60243 to −36907 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
B/op 140109 139525 −0.4% (−584) −585 to −582 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 5642 3621 −35.8% (−2021) −2021 to −2021 loss ≤ 0.05 × |baseline median| PASSED

US census dataset bufio · 17 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 231691 185485 −19.9% (−46211) −70849 to −26156 gain > 0; gain > 0.05 × |baseline median|; arm medians separated PASSED
B/op 130507 129938 −0.4% (−568) −571 to −566 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 5442 3421 −37.1% (−2021) −2021 to −2021 loss ≤ 0.05 × |baseline median| PASSED

Small empty UnmarshalBinary · 32 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 89.84 90.65 +1.05 −3.74 to +4.4 loss ≤ 0.05 × |baseline median| PASSED
B/op 48 48 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 1 1 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

Small array UnmarshalBinary · 17 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 210.2 206.1 −4.7 −17.2 to +9.6 loss ≤ 0.05 × |baseline median| PASSED
B/op 104 104 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 8 8 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

Small bitmap UnmarshalBinary · 28 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 3140 3128 −19 −117 to +74 loss ≤ 0.05 × |baseline median| PASSED
B/op 8301 8301 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 8 8 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

Small mixed UnmarshalBinary · 43 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 3552 3550 +1 −84 to +145 loss ≤ 0.05 × |baseline median| PASSED
B/op 8417 8417 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 12 12 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

Small empty file · 18 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 7205 7232 +44 −186 to +117 loss ≤ 0.05 × |baseline median| PASSED
B/op 0 0 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 0 0 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

Small empty bufio · 18 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 4428 4463 +35.5 −92 to +109 loss ≤ 0.05 × |baseline median| PASSED
B/op 0 0 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 0 0 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

Small array file · 18 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 17150 17162 +46.5 −626 to +642 loss ≤ 0.05 × |baseline median| PASSED
B/op 56 56 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 7 7 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

In-memory decode, 6500 values · 18 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 3155 3111 −19.5 −317 to +127 loss ≤ 0.05 × |baseline median| PASSED
B/op 8301 8301 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 8 8 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

In-memory decode, 65000 values · 33 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 3167 3118 −48 −215 to +148 loss ≤ 0.05 × |baseline median| PASSED
B/op 8301 8301 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 8 8 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

In-memory decode, 650000 values · 51 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 4109 4126 +10 −170 to +179 loss ≤ 0.05 × |baseline median| PASSED
B/op 8817 8817 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 27 27 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

Two bitmap containers, three-byte reads · 29 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 30974 29749 −3.6% (−1128) −3634 to −428 loss ≤ 0.05 × |baseline median| PASSED
B/op 16555 16555 0 −1 to +1 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 10 9 −10% (−1) −1 to −1 loss ≤ 0.05 × |baseline median| PASSED

Array and bitmap containers, three-byte reads · 18 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 15659 15032 −659 −2867 to +98 loss ≤ 0.05 × |baseline median| PASSED
B/op 8353 8353 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 10 10 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

Checks: 6 of 6 passed. Verification: no defect found.

Timeline