[0xaboomar]

Performance Improvements

Here are the measured performance improvements after applying the optimizations:

• Before optimization: in-use memory ranged between 70–90 MB.
• After optimization: in-use memory now ranges between 5–9 MB.

During load testing with 10K total requests and 1K concurrent requests:

• Time per request: reduced from 495 ms to 209 ms.
• Max in-use memory: reduced from 100+ MB to 16 MB.

During lower load with 1K total requests and 50 concurrent requests:

• Time per request: reduced from 25 ms to 8 ms.

---

Before

After

Note:

• These benchmarks were conducted with Watchdog monitoring two OBA servers.
• Peak points are the points where I performed load testing using Apache Bench.
• The reserved memory here is the larger virtual address space that Go pre-allocates from the OS to reduce fragmentation and minimize allocation latency.

---

What Changed

• Added lightweight StaticData and RealtimeData models containing only required GTFS fields.

• Updated StaticStore and RealtimeStore to store these minimal models instead of full bundles, reducing memory usage.

• Introduced CachedPromHandler that precomputes metrics at fixed intervals and serves cached responses.

  • Avoids expensive serialization for every scrape.

Why

• Full GTFS static/realtime bundles kept in memory were unnecessarily large.
• Prometheus scrapes were triggering full metrics serialization on every request, wasting CPU and increasing latency.

[aaronbrethorst]

lgtm! 👍