perf(replay): optimize session replay diffing — 77% faster (265→60µs) by marandaneto · Pull Request #465 · PostHog/posthog-android

marandaneto · 2026-03-19T16:17:03Z

💡 Motivation and Context

The PostHogReplayIntegration session replay engine runs a snapshot diff on every frame draw. The original diffing algorithm created 6+ intermediate collections, flattened both wireframe trees into lists, and used data class copy() for comparison — all on the hot path that executes hundreds of times per second on user devices.

This PR fundamentally redesigns the diffing algorithm and applies targeted micro-optimizations to reduce per-frame CPU and allocation overhead.

Result: 265µs → 60µs (77% faster) on a realistic 781-node wireframe tree benchmark, verified across 48 experiments.

Key optimizations

Technique	Impact
Parallel tree walk (bypass HashMap for stable structures)	-25%
`===` before `==` for nullable field comparisons	-20%
Single-pass diffing (eliminate 6+ collection ops)	-42%
Manual hex `toRGBColor` (replace `String.format`)	17× faster
Combined flatten+diff (eliminate flat list allocations)	-17%
Custom `styleEquals` (skip data class equals overhead)	-11%

Production improvements (not benchmarked)

Reusable IntArray(2) for getLocationOnScreen coordinates
Lazy MutatedNode/RemovedNode list creation (skip when empty)
Pre-sized ArrayList for events, children, mouse interactions
Simplified GradientDrawable.toRGBColor (remove unnecessary decompose/recompose)
Removed dead code (flattenChildren, findAddedAndRemovedItems from integration)

New files

RRWireframeDiffer.kt — Extracted, optimized diff engine (pure Kotlin/JVM, no Android deps)
RRWireframeDifferTest.kt — 7 correctness tests (identical trees, updates, adds, removes, structural changes, null root, flatten order, toRGBColor)
RRWireframeDifferBenchmarkTest.kt — Performance benchmark

💚 How did you test it?

7 unit tests covering all diff scenarios (added, removed, updated, structural changes, null root, flatten order, color conversion)
Benchmark test on 781-node synthetic tree (depth 4, branching factor 5, 10% changed nodes per frame)
Full :posthog:test suite passes
:posthog-android:compileDebugKotlin compiles cleanly
48 experiment iterations with automated correctness checks

📝 Checklist

I reviewed the submitted code.
I added tests to verify the changes.
I updated the docs if needed.
No breaking change or entry added to the changelog.

…ask on 781-node tree Result: {"status":"keep","total_µs":265,"color_µs":317,"mask_µs":42,"node_count":781,"p95_µs":307}

…nChildren, manual hex toRGBColor. 265→153µs (-42%), color 317→18µs (-94%) Result: {"status":"keep","total_µs":153,"color_µs":18,"mask_µs":39,"node_count":781,"p95_µs":216}

…ems, remaining = removed. 153→131µs (-14%) Result: {"status":"keep","total_µs":131,"color_µs":19,"mask_µs":36,"node_count":781,"p95_µs":163}

… Avoids stack object management. 131→119µs (-9%) Result: {"status":"keep","total_µs":119,"color_µs":16,"mask_µs":26,"node_count":781,"p95_µs":134}

…ashMap, avoids allocating two 781-element flat lists. 119→99µs (-17%), total from baseline 265→99µs (-63%) Result: {"status":"keep","total_µs":99,"color_µs":19,"mask_µs":48,"node_count":781,"p95_µs":117}

…tegration.kt. No benchmark change (same code), but production integration now uses optimized paths. Result: {"status":"keep","total_µs":107,"color_µs":18,"mask_µs":26,"node_count":781,"p95_µs":119}

…quality check before deep equals. Minimal bench impact but helps in production where styles are often same object. Result: {"status":"keep","total_µs":106,"color_µs":17,"mask_µs":46,"node_count":781,"p95_µs":125}

…om PostHogReplayIntegration (replaced by RRWireframeDiffer). Code cleanup, no perf change. Result: {"status":"keep","total_µs":106,"color_µs":17,"mask_µs":46,"node_count":781,"p95_µs":125}

…y without HashMap when IDs match. Falls back to HashMap only for structural changes. 106→80µs (-25%) Result: {"status":"keep","total_µs":80,"color_µs":19,"mask_µs":47,"node_count":781,"p95_µs":84}

…(skip when empty), pre-sized ArrayLists, single-root diffTrees overload. No bench change but reduces per-frame allocations in production. Result: {"status":"keep","total_µs":81,"color_µs":17,"mask_µs":41,"node_count":781,"p95_µs":89}

…e — avoids per-view allocation. Production-only improvement, no bench impact. Result: {"status":"keep","total_µs":81,"color_µs":17,"mask_µs":41,"node_count":781,"p95_µs":89}

…yList when ViewGroup has children, pre-sized to childCount. Saves 1 allocation per leaf view. Result: {"status":"keep","total_µs":81,"color_µs":17,"mask_µs":41,"node_count":781,"p95_µs":89}

…Minor allocation reduction in production. Result: {"status":"keep","total_µs":81,"color_µs":17,"mask_µs":41,"node_count":781,"p95_µs":89}

…directly, avoids unnecessary list wrapping and parallel walk setup. Result: {"status":"keep","total_µs":81,"color_µs":18,"mask_µs":42,"node_count":781,"p95_µs":112}

…70% faster). Optimization space exhausted. Result: {"status":"keep","total_µs":80,"color_µs":14,"mask_µs":23,"node_count":781,"p95_µs":89}

…een/blue decomposition + Color.rgb recomposition (toRGBColor already masks to 0xFFFFFF). Fix autoresearch.sh stale test cache. Result: {"status":"keep","total_µs":81,"color_µs":17,"mask_µs":51,"node_count":781,"p95_µs":86}

…, updates, adds, removes, structural changes, null root, flattenChildren, toRGBColor). All pass. 79µs confirmed. Result: {"status":"keep","total_µs":79,"color_µs":15,"mask_µs":24,"node_count":781,"p95_µs":87}

…en oldList.size == newList.size (the common case). 80→78µs (~3%) Result: {"status":"keep","total_µs":78,"color_µs":17,"mask_µs":30,"node_count":781,"p95_µs":87}

… 78-80µs. Optimization space fully explored — 33 experiments, 71% total improvement. Result: {"status":"keep","total_µs":80,"color_µs":16,"mask_µs":43,"node_count":781,"p95_µs":86}

…ields in wireframePropertiesEqual. Handles null-null fast and interned strings. 80→73µs (-8.7%) Result: {"status":"keep","total_µs":73,"color_µs":16,"mask_µs":42,"node_count":781,"p95_µs":89}

…yle. Avoids Intrinsics.areEqual for null-null comparisons (most fields are null). 73→71µs (-3%) Result: {"status":"keep","total_µs":71,"color_µs":17,"mask_µs":43,"node_count":781,"p95_µs":82}

…ontSize, padding, etc). null===null skips Intrinsics.areEqual function call. 71→63µs (-11%) Result: {"status":"keep","total_µs":63,"color_µs":18,"mask_µs":36,"node_count":781,"p95_µs":70}

…esEqual (disabled, checked, max, parentId). 63→62µs (~1.5%) Result: {"status":"keep","total_µs":62,"color_µs":16,"mask_µs":27,"node_count":781,"p95_µs":69}

Result: {"status":"keep","total_µs":60,"color_µs":14,"mask_µs":43,"node_count":781,"p95_µs":66}

…rectness tests pass, Android module compiles. Result: {"status":"keep","total_µs":61,"color_µs":19,"mask_µs":43,"node_count":781,"p95_µs":65}

…. Optimization surface exhausted — remaining ideas require Android runtime. Result: {"status":"keep","total_µs":59,"color_µs":17,"mask_µs":38,"node_count":781,"p95_µs":68}

…ode, same performance. JIT handles both paths equally. Result: {"status":"keep","total_µs":61,"color_µs":16,"mask_µs":25,"node_count":781,"p95_µs":68}

…(77%). Optimization surface fully exhausted on JVM benchmark. Remaining ideas are Android-only. Result: {"status":"keep","total_µs":61,"color_µs":20,"mask_µs":24,"node_count":781,"p95_µs":70}

github-advanced-security

CodeQL found more than 20 potential problems in the proposed changes. Check the Files changed tab for more details.

github-actions · 2026-03-19T16:25:46Z

posthog-android Compliance Report

Date: 2026-03-19 16:25:45 UTC
Duration: 147109ms

⚠️ Some Tests Failed

28/29 tests passed, 1 failed

Capture Tests

⚠️ 28/29 tests passed, 1 failed

View Details

Test	Status	Duration
Format Validation.Event Has Required Fields	✅	2320ms
Format Validation.Event Has Uuid	✅	2027ms
Format Validation.Event Has Lib Properties	✅	2027ms
Format Validation.Distinct Id Is String	✅	2026ms
Format Validation.Token Is Present	✅	2023ms
Format Validation.Custom Properties Preserved	✅	2023ms
Format Validation.Event Has Timestamp	✅	2026ms
Retry Behavior.Retries On 503	✅	7025ms
Retry Behavior.Does Not Retry On 400	✅	4025ms
Retry Behavior.Does Not Retry On 401	✅	4026ms
Retry Behavior.Respects Retry After Header	✅	7026ms
Retry Behavior.Implements Backoff	✅	17030ms
Retry Behavior.Retries On 500	✅	7026ms
Retry Behavior.Retries On 502	✅	7025ms
Retry Behavior.Retries On 504	✅	7023ms
Retry Behavior.Max Retries Respected	✅	17034ms
Deduplication.Generates Unique Uuids	✅	2043ms
Deduplication.Preserves Uuid On Retry	✅	7026ms
Deduplication.Preserves Uuid And Timestamp On Retry	✅	12022ms
Deduplication.Preserves Uuid And Timestamp On Batch Retry	✅	7028ms
Deduplication.No Duplicate Events In Batch	✅	2032ms
Deduplication.Different Events Have Different Uuids	✅	2022ms
Compression.Sends Gzip When Enabled	✅	2017ms
Batch Format.Uses Proper Batch Structure	✅	2017ms
Batch Format.Flush With No Events Sends Nothing	✅	2015ms
Batch Format.Multiple Events Batched Together	✅	2030ms
Error Handling.Does Not Retry On 403	✅	4020ms
Error Handling.Does Not Retry On 413	❌	4022ms
Error Handling.Retries On 408	✅	7022ms

Failures

error_handling.does_not_retry_on_413

Expected 1 requests, got 2

marandaneto added 30 commits March 19, 2026 13:05

autoresearch: setup benchmark for PostHogReplayIntegration perf

99d213d

fix: spotless formatting

202fc72

Baseline: flattenChildren + findAddedAndRemovedItems + toRGBColor + m…

8088689

…ask on 781-node tree Result: {"status":"keep","total_µs":265,"color_µs":317,"mask_µs":42,"node_count":781,"p95_µs":307}

Single-pass diffing (eliminate 6+ collection ops), stack-based flatte…

879fc2e

…nChildren, manual hex toRGBColor. 265→153µs (-42%), color 317→18µs (-94%) Result: {"status":"keep","total_µs":153,"color_µs":18,"mask_µs":39,"node_count":781,"p95_µs":216}

Eliminate seenOldIds HashSet: use oldMap.remove() to track matched it…

136e79c

…ems, remaining = removed. 153→131µs (-14%) Result: {"status":"keep","total_µs":131,"color_µs":19,"mask_µs":36,"node_count":781,"p95_µs":163}

Recursive flattenInto with single result list instead of stack-based.…

6af00b9

… Avoids stack object management. 131→119µs (-9%) Result: {"status":"keep","total_µs":119,"color_µs":16,"mask_µs":26,"node_count":781,"p95_µs":134}

Wire up RRWireframeDiffer.diffTrees and toRGBColor in PostHogReplayIn…

5649833

…tegration.kt. No benchmark change (same code), but production integration now uses optimized paths. Result: {"status":"keep","total_µs":107,"color_µs":18,"mask_µs":26,"node_count":781,"p95_µs":119}

Remove dead code: old flattenChildren and findAddedAndRemovedItems fr…

0e09d5e

…om PostHogReplayIntegration (replaced by RRWireframeDiffer). Code cleanup, no perf change. Result: {"status":"keep","total_µs":106,"color_µs":17,"mask_µs":46,"node_count":781,"p95_µs":125}

Parallel tree walk diffTrees: compare nodes at same positions directl…

49a51b0

…y without HashMap when IDs match. Falls back to HashMap only for structural changes. 106→80µs (-25%) Result: {"status":"keep","total_µs":80,"color_µs":19,"mask_µs":47,"node_count":781,"p95_µs":84}

Reuse IntArray(2) for coordinates in toWireframe/toScreenshotWirefram…

bda429d

…e — avoids per-view allocation. Production-only improvement, no bench impact. Result: {"status":"keep","total_µs":81,"color_µs":17,"mask_µs":41,"node_count":781,"p95_µs":89}

Avoid empty list allocation in toWireframe: only create children Arra…

f6e3f3a

…yList when ViewGroup has children, pre-sized to childCount. Saves 1 allocation per leaf view. Result: {"status":"keep","total_µs":81,"color_µs":17,"mask_µs":41,"node_count":781,"p95_µs":89}

Pre-size ArrayList for events (3), mouseInteractions (pointerCount). …

7d749df

…Minor allocation reduction in production. Result: {"status":"keep","total_µs":81,"color_µs":17,"mask_µs":41,"node_count":781,"p95_µs":89}

Optimize single-root diffTrees: handle null oldRoot (first snapshot) …

8953b8b

…directly, avoids unnecessary list wrapping and parallel walk setup. Result: {"status":"keep","total_µs":81,"color_µs":18,"mask_µs":42,"node_count":781,"p95_µs":112}

Final verification run. Stable at 80µs. Total improvement: 265→80µs (…

b25dcfd

…70% faster). Optimization space exhausted. Result: {"status":"keep","total_µs":80,"color_µs":14,"mask_µs":23,"node_count":781,"p95_µs":89}

Same-size fast path in parallelWalk: skip minOf and trailing loops wh…

172d809

…en oldList.size == newList.size (the common case). 80→78µs (~3%) Result: {"status":"keep","total_µs":78,"color_µs":17,"mask_µs":30,"node_count":781,"p95_µs":87}

Update autoresearch docs, checks include correctness tests. Stable at…

c6aa5b9

… 78-80µs. Optimization space fully explored — 33 experiments, 71% total improvement. Result: {"status":"keep","total_µs":80,"color_µs":16,"mask_µs":43,"node_count":781,"p95_µs":86}

Reference equality (===) before deep equality (==) for all nullable f…

aeb93fd

…ields in wireframePropertiesEqual. Handles null-null fast and interned strings. 80→73µs (-8.7%) Result: {"status":"keep","total_µs":73,"color_µs":16,"mask_µs":42,"node_count":781,"p95_µs":89}

Custom styleEquals with === for all 10 nullable string fields in RRSt…

485d160

…yle. Avoids Intrinsics.areEqual for null-null comparisons (most fields are null). 73→71µs (-3%) Result: {"status":"keep","total_µs":71,"color_µs":17,"mask_µs":43,"node_count":781,"p95_µs":82}

Apply === fast path to Int? fields in styleEquals too (borderWidth, f…

a6d01ac

…ontSize, padding, etc). null===null skips Intrinsics.areEqual function call. 71→63µs (-11%) Result: {"status":"keep","total_µs":63,"color_µs":18,"mask_µs":36,"node_count":781,"p95_µs":70}

Apply === fast path to remaining nullable fields in wireframeProperti…

546d65c

…esEqual (disabled, checked, max, parentId). 63→62µs (~1.5%) Result: {"status":"keep","total_µs":62,"color_µs":16,"mask_µs":27,"node_count":781,"p95_µs":69}

Doc updates. Continued verification: 60µs. 265→60µs (77% improvement).

0640072

Result: {"status":"keep","total_µs":60,"color_µs":14,"mask_µs":43,"node_count":781,"p95_µs":66}

Final verification. Stable at 60-62µs. Total: 265→61µs (77%). All cor…

21d49ae

…rectness tests pass, Android module compiles. Result: {"status":"keep","total_µs":61,"color_µs":19,"mask_µs":43,"node_count":781,"p95_µs":65}

Continued verification: 59µs. Best so far. 265→59µs (78% improvement)…

b7fb79d

…. Optimization surface exhausted — remaining ideas require Android runtime. Result: {"status":"keep","total_µs":59,"color_µs":17,"mask_µs":38,"node_count":781,"p95_µs":68}

Remove parallelWalk code duplication (same-size fast path). Simpler c…

8671d4e

…ode, same performance. JIT handles both paths equally. Result: {"status":"keep","total_µs":61,"color_µs":16,"mask_µs":25,"node_count":781,"p95_µs":68}

Final session — update ideas backlog. 45 experiments total, 265→60µs …

01d3a14

…(77%). Optimization surface fully exhausted on JVM benchmark. Remaining ideas are Android-only. Result: {"status":"keep","total_µs":61,"color_µs":20,"mask_µs":24,"node_count":781,"p95_µs":70}

Final quality gate: all tests pass, 265→60µs (77%). 48 experiments.

6358306

github-advanced-security bot found potential problems Mar 19, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

perf(replay): optimize session replay diffing — 77% faster (265→60µs)#465

perf(replay): optimize session replay diffing — 77% faster (265→60µs)#465
marandaneto wants to merge 31 commits intomainfrom
autoresearch/replay-perf-20260319

marandaneto commented Mar 19, 2026

Uh oh!

github-advanced-security bot left a comment

Uh oh!

github-actions bot commented Mar 19, 2026

Failures

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

marandaneto commented Mar 19, 2026

💡 Motivation and Context

Key optimizations

Production improvements (not benchmarked)

New files

💚 How did you test it?

📝 Checklist

Uh oh!

github-advanced-security bot left a comment

Choose a reason for hiding this comment

Uh oh!

github-actions bot commented Mar 19, 2026

posthog-android Compliance Report

⚠️ Some Tests Failed

Capture Tests

Failures

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants