Storine v0.3.0

Reference

Known limitations

What Storine does not do yet, and where it trails other databases.

  • Tail latency trails RocksDB in one case: 4 writers overwriting on 4 CPUs, p99.99 2.2–2.5 ms against 0.26–0.30 ms. That is the scheduler: four writers plus the flusher are more busy threads than CPUs. With a CPU to spare, it is about 70 µs. Bulk loads match RocksDB when the store is told its size (expected_records). Durable p99.9 and p99.99 trail LMDB on fast-sync disks (benchmark report, sections 8.5 and 8.11).
  • Single-group cluster writes wait for two disk syncs in a row, the leader's log and then a follower's: about 8 ms on disks whose syncs take 4 ms. Concurrency raises throughput (section 8.8).
  • The server does not expose the cold and paged index tiers. They are in the embedded API.
  • Point-in-time recovery is not available on range clusters, which have cluster backups instead.
  • Changing a node's locality in a range cluster takes a restart of every node.
  • Range clusters list only collections with an index or metadata.
  • TLS supports one cipher suite per version, Ed25519 certificates only, and no client certificates. .NET on Windows cannot connect with TLS, as its TLS stack lacks Ed25519.
  • Change feeds exist for range clusters only, through storage-cli changefeed. The client libraries do not expose them.
  • The local storage-cli data commands work on the typed demo store, not on a server's document data.
  • Upgrades between minor versions before 1.0 need every node of a cluster stopped, and a data directory written by a newer version cannot be opened by an older one (upgrades).
  • Client libraries are installed from the repository; they are not on package registries yet.
  • Testing: soaks ran for 2 hours per cluster mode under constant faults; no multi-day soak has run. Performance figures and soaks are from Linux; macOS and Windows pass the test suite in CI and smoke tests in the release build.
  • The SQL gateway has limitations of its own, listed in its documentation.