Leveraging Prometheus’ TSDB for conprof (Continuous Profiling)
summary
conprof reuses Prometheus components (service discovery, scrape manager, remote write) and modifies TSDB to store profiles as timestamp + byte slices, split timestamp and value chunks so the UI can read timestamps without decompressing profiles. Samples are grouped (currently 12-sample groups) and compressed with zstd, yielding roughly 50–75% disk savings at a runtime overhead of about 0.2–3%. A basic UI lets users query by Prometheus-style labels and fetch individual pprof heap/cpu profiles on demand; storage and query optimizations continue to evolve.
AI-written summary · about
chapters (9)
- 0:00Start
- 0:17About Me
- 1:21The Difference of Profiling and Metrics
- 2:00Cpu Throttling High Alert
- 8:10Perf Data Converter
- 13:28Demo
- 15:29Flame Graph
- 17:24The Control of Time Series Database
- 17:54Tsdb Appender Interface
