Parquet Blocks Production Readiness

[annanay25]

Meta issue to track production readiness of the Apache Parquet block format.

References

• Design proposal #1407
• Initial PR adding parquet support #1479

Tracking production readiness

• [ ] Fix FindTraceByID [@mdisibio]
  • [x] Fix panic when scanning rowgroups with multiple pages
  • [x] Current bug with find by ID logic? (https://github.com/grafana/tempo/pull/1531)
• [x] Serverless build
  • [x] Make a decision on Google Cloud Run vs. Cloud Functions
  • [x] Swap serverless querying to vParquet
  • [x] Rework our CI pipeline
• [ ] Caching [@annanay25]
  • [ ] Put metadata/footer in real cache.
• [ ] Fine tuning [@joe-elliott]
  • [ ] Row group size (100MB? 10000 traces?)
  • [ ] Read buffer size/count (1MB x 16 currently)
  • [ ] Search shard target bytes per job (30MB, 100MB?)
  • [ ] Sort inner trace content: Sort batches by span name - does this significantly improve the common case of searching for span by name?
• [ ] Compaction [@mdisibio]
  • [x] Pre-fetch before combining?
  • [ ] Fix combiner metrics -- move combiner into versioned encodings
• [ ] Performance/Stability testing in highly multitenant environments
• [ ] Upstream issues: [@mdisibio]
  • [x] https://github.com/segmentio/parquet-go/issues/250
  • [x] https://github.com/segmentio/parquet-go/issues/254

Nice to haves (non-blocking)

• [ ] CLI (prints schema & column sizes already) [@mdisibio]
  • [x] iterate through a column and debug print all values https://github.com/grafana/tempo/pull/1531

Future

The following are out of scope for production readiness of parquet blocks, rather its a rough roadmap for the parquet block format.

• [ ] Cache specific resource attribute columns? cluster / namespace / service.name ..
• [ ] Switch WAL over to vParquet
• [ ] Dynamic column support
• [ ] Implement TraceQL over Parquet blocks.