Browser pool: pre-warm Chromium instances to eliminate cold-start latency on search #47

New issue

Closed

opened 2026-04-20 08:50:40 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-04-20 08:50:40 -07:00

Owner

Problem

Each /api/search request cold-starts a new Chromium instance to bypass Kasada. Under load this means 10-30s of browser startup before any eBay content is fetched. Spike #5 showed p50 of 29s on search — startup overhead is the dominant cost.

Proposed Fix

Maintain a pool of N pre-warmed Chromium/Playwright browser contexts (suggested N=2-3). Each search request borrows a context from the pool, navigates and scrapes, then returns the context (not closes it).

Pool initialized at startup, replenished after each use
Idle timeout to close contexts if unused for >5 min (avoid memory bloat)
Pool size configurable via BROWSER_POOL_SIZE env var (default 2)
Fallback: spin up a fresh browser if pool is empty (current behavior)

Expected Impact

p50 search latency: 29s → ~3-5s (Kasada negotiation only, no cold start)
Enables meaningful concurrency without N×browser memory overhead

Notes

Playwright browser contexts are lightweight vs full browser instances — pool the context, not the browser
Kasada state may need to be preserved per-context (cookies, fingerprint) — test whether context reuse maintains bypass

## Problem Each `/api/search` request cold-starts a new Chromium instance to bypass Kasada. Under load this means 10-30s of browser startup before any eBay content is fetched. Spike #5 showed p50 of 29s on search — startup overhead is the dominant cost. ## Proposed Fix Maintain a pool of N pre-warmed Chromium/Playwright browser contexts (suggested N=2-3). Each search request borrows a context from the pool, navigates and scrapes, then returns the context (not closes it). - Pool initialized at startup, replenished after each use - Idle timeout to close contexts if unused for >5 min (avoid memory bloat) - Pool size configurable via `BROWSER_POOL_SIZE` env var (default 2) - Fallback: spin up a fresh browser if pool is empty (current behavior) ## Expected Impact - p50 search latency: 29s → ~3-5s (Kasada negotiation only, no cold start) - Enables meaningful concurrency without N×browser memory overhead ## Notes - Playwright browser contexts are lightweight vs full browser instances — pool the context, not the browser - Kasada state may need to be preserved per-context (cookies, fingerprint) — test whether context reuse maintains bypass