diff options
| -rw-r--r-- | DOCS/benchmarks.md | 113 |
1 files changed, 66 insertions, 47 deletions
diff --git a/DOCS/benchmarks.md b/DOCS/benchmarks.md index 56e903a..96b4fbb 100644 --- a/DOCS/benchmarks.md +++ b/DOCS/benchmarks.md @@ -14,9 +14,8 @@ ### Summary The `make check` workflow consists of: 1. `golangci-lint run` (Backend) -2. `npm run lint` (Frontend) -3. `go test -cover ./...` (Backend) -4. `npm test -- --run` (Frontend) +2. `go test -cover ./...` (Backend) +3. `npm test -- --run` (Frontend - frontend-vanilla) The goal of keeping the check under 15 seconds for a fast local feedback loop has been achieved. @@ -24,65 +23,75 @@ The goal of keeping the check under 15 seconds for a fast local feedback loop ha ## Go Backend Benchmarks -**Environment:** Linux arm64, Go, SQLite -**Date:** 2026-02-16 -**Methodology:** `go test -bench=. -benchmem -count=3` +**Environment:** Linux amd64, Intel(R) Xeon(R) Platinum 8581C @ 2.10GHz, Go 1.24, SQLite +**Date:** 2026-02-17 +**Methodology:** `go test -bench=. -benchmem -count=3 -run='^$'` ### API Handlers (`api/`) -| Benchmark | ops | ns/op | B/op | allocs/op | -|---|---|---|---|---| -| HandleStream | 6,664 | 168,344 | 380,350 | 1,423 | -| HandleStreamWithSearch | 6,200 | 192,454 | 381,033 | 1,432 | -| HandleItemUpdate | 24,469 | 48,513 | 8,597 | 46 | -| HandleFeedList | 51,320 | 22,846 | 10,308 | 117 | +| Benchmark | ns/op | B/op | allocs/op | +|---|---|---|---| +| HandleStream | 780,000 | 379,930 | 1,419 | +| HandleStreamWithSearch | 942,618 | 380,511 | 1,428 | +| HandleItemUpdate | 6,228,144 | 8,556 | 46 | +| HandleFeedList | 295,920 | 10,327 | 117 | -**Findings:** Stream endpoints (~170us) are dominated by SQLite query + JSON serialization. Search adds ~14% overhead via FTS. Item updates and feed listing are fast (~48us and ~23us respectively). The ~380KB/op allocation for stream is from serializing item content; this could be reduced by excluding `full_content` in list views. +**Findings:** Stream endpoints (~780µs) are dominated by SQLite query + JSON serialization. Search adds ~21% overhead via FTS. Item updates (~6.2ms) include a full DB write cycle. Feed listing is fast (~296µs). The ~380KB/op allocation for stream comes from serializing item descriptions; the `full_content` field is already excluded from list views (see item model benchmarks for full_content savings data). ### Crawler (`internal/crawler/`) -| Benchmark | ops | ns/op | B/op | allocs/op | -|---|---|---|---|---| -| ParseFeed | 12,157 | 98,513 | 92,216 | 1,643 | -| CrawlFeedMocked | 1,497 | 782,713 | 169,491 | 2,233 | -| GetFeedContent | 9,720 | 122,448 | 46,986 | 190 | +| Benchmark | ns/op | B/op | allocs/op | +|---|---|---|---| +| ParseFeed | 220,619 | 92,292 | 1,643 | +| CrawlFeedMocked | 3,034,479 | 170,857 | 2,236 | +| GetFeedContent | 1,150,492 | 46,547 | 190 | -**Findings:** Feed parsing (~98us) is fast. Full crawl cycle (~783us mocked) is dominated by HTTP round-trip + DB write. Content fetching (~122us) includes HTTP + HTML sanitization. +**Findings:** Feed parsing (~221µs) is fast. Full crawl cycle (~3ms mocked) is dominated by HTTP round-trip + DB write. Content fetching (~1.15ms) includes HTTP + HTML sanitization. ### Item Model (`models/item/`) -| Benchmark | ops | ns/op | B/op | allocs/op | -|---|---|---|---|---| -| ItemCreate | 21,397 | 55,924 | 1,415 | 22 | -| ItemCreateBatch100 | 216 | 5,574,677 | 139,213 | 2,100 | -| Filter_Empty | 62,834 | 19,005 | 13,096 | 82 | -| Filter_15Items | 7,429 | 159,417 | 373,392 | 1,771 | -| Filter_WithFTS | 6,406 | 184,237 | 374,120 | 1,779 | -| Filter_WithImageProxy | 5,348 | 216,287 | 496,533 | 2,491 | -| FilterPolicy | 104,919 | 11,444 | 17,768 | 150 | -| RewriteImages | 182,242 | 6,453 | 11,048 | 97 | -| ItemSave | 28,552 | 41,641 | 592 | 11 | -| Filter_LargeDataset | 8,623 | 139,067 | 361,769 | 1,186 | - -**Findings:** Image proxy adds ~35% overhead to filtering (216us vs 159us) due to URL rewriting of `<img>` tags. Batch inserts scale linearly (~56us/item). The `FilterPolicy` HTML sanitizer is fast at ~11us. Full-text search adds minimal overhead (~15%) to filtering. +| Benchmark | ns/op | B/op | allocs/op | +|---|---|---|---| +| ItemCreate | 6,816,479 | 1,388 | 21 | +| ItemCreateBatch100 | 717,721,761 | 138,539 | 2,102 | +| Filter_Empty | 291,475 | 13,209 | 82 | +| Filter_15Items | 799,494 | 373,444 | 1,767 | +| Filter_WithFTS | 850,579 | 374,073 | 1,775 | +| Filter_WithImageProxy | 1,014,302 | 496,582 | 2,487 | +| FilterPolicy | 25,979 | 17,768 | 150 | +| RewriteImages | 17,331 | 11,048 | 97 | +| ItemSave | 6,076,841 | 592 | 11 | +| Filter_LargeDataset | 771,772 | 361,818 | 1,182 | +| **Filter_15Items_WithFullContent** | **798,356** | **362,778** | **1,167** | +| **Filter_15Items_IncludeFullContent** | **1,322,135** | **594,842** | **4,054** | +| **Filter_LargeDataset_WithFullContent** | **772,676** | **363,001** | **1,182** | + +**Key Finding – full_content exclusion savings:** +When items have scraped `full_content` (~2KB each), excluding it from list responses (default behavior) vs including it: +- **Memory**: 363KB/op vs 595KB/op — **39% reduction** in allocations +- **Speed**: 798µs vs 1,322µs — **40% faster** + +This confirms the value of the full_content exclusion from list views (implemented in NK-k9otuy). The `Filter_LargeDataset_WithFullContent` benchmark (500 items with full_content, excluded from response) shows only 363KB/op — nearly identical to the no-content baseline — demonstrating that the exclusion scales well. + +**Other Findings:** Image proxy adds ~27% overhead to filtering (1014µs vs 799µs) due to URL rewriting of `<img>` tags. Batch inserts scale linearly. HTML sanitizer (`FilterPolicy`) is fast at ~26µs. ### Web Middleware (`web/`) -| Benchmark | ops | ns/op | B/op | allocs/op | -|---|---|---|---|---| -| GzipMiddleware | 100,623 | 11,881 | 11,999 | 25 | -| SecurityHeaders | 484,862 | 2,402 | 6,185 | 22 | -| CSRFMiddleware | 495,804 | 2,395 | 6,028 | 23 | -| FullMiddlewareStack | 362,329 | 3,237 | 8,745 | 34 | +| Benchmark | ns/op | B/op | allocs/op | +|---|---|---|---| +| GzipMiddleware | 22,287 | 12,945 | 25 | +| SecurityHeaders | 7,081 | 6,190 | 22 | +| CSRFMiddleware | 7,250 | 6,037 | 23 | +| FullMiddlewareStack | 14,886 | 9,377 | 34 | -**Findings:** The full middleware stack adds only ~3.2us per request. Gzip compression is the most expensive middleware (~12us) due to compression work, but is only applied to compressible responses. CSRF and security headers are near-zero cost (~2.4us each). +**Findings:** The full middleware stack adds only ~15µs per request. Gzip compression is the most expensive middleware (~22µs) due to compression work, but is only applied to compressible responses. CSRF and security headers are near-zero cost (~7µs each). --- ## Frontend Performance Tests (Vanilla JS / v3) -**Environment:** Vitest + jsdom -**Date:** 2026-02-16 +**Environment:** Vitest + jsdom, Node.js 20 +**Date:** 2026-02-17 ### Store Operations @@ -103,18 +112,28 @@ The goal of keeping the check under 15 seconds for a fast local feedback loop ha | createFeedItem (500 items) | < 200ms | PASS | | createFeedItem (1000 items) | < 100ms | PASS | | DOM insertion (100 items) | < 200ms | PASS | -| DOM insertion (500 items) | < 500ms | PASS | +| DOM insertion (500 items) | < 500ms | PASS (~324ms) | -**Findings:** All frontend performance tests pass well within their thresholds. The vanilla JS approach with direct DOM manipulation and simple event emitter pattern keeps operations fast. Store updates with 500+ items and event dispatch remain under 10ms. +**Findings:** All frontend performance tests pass well within their thresholds. The vanilla JS approach with direct DOM manipulation and simple event emitter pattern keeps operations fast. Store updates with 500+ items and event dispatch remain under 10ms. DOM insertion of 500 items takes ~324ms, comfortably within the 500ms threshold. + +--- + +## Notable Changes Since Last Benchmark (2026-02-16) + +1. **New full_content benchmarks** (NK-ekxfvv): Three new benchmarks quantify the memory savings from excluding `full_content` in list responses. The 39% memory reduction and 40% speed improvement are now measured with realistic ~2KB article content. + +2. **safehttp SSRF fix**: HTTP proxy bypass bug fixed — safe client now uses `Proxy: nil` to prevent proxy environment variables from bypassing the private IP check. + +3. **Environment**: These benchmarks were run on amd64 Intel Xeon @ 2.10GHz vs previous arm64. Raw times are not directly comparable between runs due to architecture differences. --- ## Potential Improvements -1. **Stream endpoint allocations**: The ~380KB/op for stream could be reduced by excluding `full_content` from list views and only fetching it on demand (already partially implemented via the scrape endpoint). +1. **Stream endpoint allocations**: The ~380KB/op for stream comes from description serialization. With full_content properly excluded, the main remaining allocation is in description strings. Further reduction possible with response streaming. -2. **Image proxy overhead**: The 35% filtering overhead from image rewriting could be cached or deferred to the client side. +2. **Image proxy overhead**: The ~27% filtering overhead from image rewriting could be cached or pre-processed at ingest time. 3. **Batch operations**: The item batch insert benchmark shows good linear scaling; could be leveraged for bulk import operations. -4. **Gzip middleware**: At ~12us, it's the most expensive middleware. Consider pre-compressing static assets and only applying runtime gzip to API responses. +4. **Gzip middleware**: At ~22µs, it's the most expensive middleware. Consider pre-compressing static assets and only applying runtime gzip to API responses. |