Armada KVArmada KV
/vmain
Documentation
Edit on GitHub

Changelog

v0.5.2 (unreleased)

Bugfixes

  • Fix regatta.v1.KV/IterateRange without range_end returning 0 results.

v0.5.1

Improvements

  • API server will now by default use fixed number of pre-created worker go routines for stream handling. The behaviour could be overridden by setting api.stream-workers config value.
  • regatta.v1.KV/IterateRange now uses coroutine instead of full goroutine.

Security

  • API server max number of concurrent connections can be set by api.max-concurrent-connections config value.
  • API server max number of concurrent streams can be set by api.max-concurrent-streams config value.

Bugfixes

  • Fix storage compaction metrics.

v0.5.0

Highlights

Follower to leader replication

  • The write API (Put, DeleteRange, Txn) is now available on leader cluster nodes as well.
  • The API provides read after write guarantee by relying on backpropagation from the leader.

Dynamic tables management

  • Tables now could be managed dynamically during the runtime of the server using newly provided regatta.v1.Tables API.
  • Tables API could be secured by an API token using tables.token configuration value.

Security

  • API port now supports mTLS
  • API port mTLS can verify Hostname or CN in presented certificate.
  • Follower to Leader mTLS can now verify Hostname or CN in presented certificate.

Improvements

  • Improve on API allocations when deserializing gRPC messages.
  • Support zstd API compression.
  • Removed obsolete table manager cache that could have caused hard to debug race conditions.
  • Bump to Go 1.22.

Deprecations

  • tables.names and tables.delete configuration values were deprecated and will be removed in future releases.

Bugfixes

  • Proper authentication of maintenance.v1.Backup/Restore API endpoint.
  • When table is deleted in the leader cluster the followers will gracefully handle the situation by deleting the table locally and stopping the replication.
  • Fixed potential Engine deadlock when shutting down the server.

v0.4.1

Improvements

  • Server now reports transient errors for requests that could be potentially retried as codes.Unavailable.

Bugfixes

  • Fixed the default timeout of KV/IterateRange operation.

v0.4.0

Breaking changes

  • Remove maintenance server port, the API is now available on standard API port. It could still though be secured by a separate API token.
  • Replaced clusterID with shard_id in metric label names.

Features

  • Added new regatta.v1.KV/IterateRange API.
  • Added option to get the server configuration via the regatta.v1.Cluster/Status API.
  • Added memberlist.cluster-name and memberlist.node-name options.
    • Both help identifying cluster and nodes in CLuster API responses.
    • Cluster name setting prevents accidental pairing of multiple clusters running in the same network. (with different cluster name setting)
  • Added api.advertise-address option to split address advertised to clients from the listen address.
  • Removed maintenance.address config option.
  • Removed maintenance.cert-filename config option.
  • Removed maintenance.key-filename config option.

Bugfixes

  • Fix KV/DeleteRange Count erroneously returning 1 if no key was deleted.
  • Fix KV/DeleteRange PrevKV panicking when previous key did not exist.

v0.3.2

A bugfix release, fixes critical issue in replication server TLS handling.

Bugfixes

  • Fix replication server TLS.

v0.3.1

A bugfix release, with small improvements.

Improvements

  • Improve startup logging and failure reporting.

Bugfixes

  • Fix rest.address handling.
  • Fix race-conditions in Raft events handling.

v0.3.0

Release brings a brand-new Cluster API as well as ability to listen over plain or unix socket connection.

Breaking changes

  • Semantics of api.address replication.address and maintenance.address changed. The protocol is now mandatory could be one of http|https|unix|unixs.
  • Removed api.reflection-api configuration.

Features

  • Armada could now serve its api on unix sockets.
  • Support for serving over plain (non-TLS) gRPC added.
  • Added regatta.v1.Cluster api.

Improvements

  • Improve version command output.
  • Follower can connect to leader over plain connection or via unix socket.

Bugfixes

  • Fix cluster member node registration.

v0.2.3

Release with profile guided optimization.

Improvements

  • Removal of few obsolete dependencies.
  • Set toolchain in go.mod file.
  • Enable profile guided optimizations.

v0.2.2

Minor release with just updated dependencies.

v0.2.1

Release focuses on stability and performance of the leader-follower replication.

Breaking changes

  • Leader replication.logCacheSize now defaults to 0.

Features

  • Add replication.keepalive-time config option for follower. Sets how often the keepalive should be sent.
  • Add replication.keepalive-timeout config option for follower. Sets how long to wait for an ack of the keepalive message.

Improvements

  • Improve follower replication throughput and latency.
  • Snapshot recovery type could be switched on a running cluster safely.
  • Bump to Go 1.21.
  • Keepalive for replication connection to tackle misbehaving LBs.

Bugfixes

  • Fix release image cosign.
  • Fix off by one replication error.

v0.2.0

Release is mostly focused on tidying up the docs and code base and on resource consumption improvements.

  • New way of producing table snapshots is introduced that should lead to quicker and more efficient process of recovering shards in a Raft cluster.
  • Experimental Tan logdb feature was promoted to a regular and made a default choice.
  • Gossip based cluster registry was added, the necessary migration is handled automatically by the new version of the armadakv/armada-helm helm chart. If you want to know more refer to the documentation of memberlist flags.

Breaking changes

  • Removal of --experimental.tanlogdb flag.
  • Addition of gossip based cluster registry.
  • regatta.v1.KV/Range now returns empty response instead of NOT_FOUND error if result of the query is an empty range. Implements proposal #001.

Features

  • New --raft.logdb flag (defaults to Tan).
  • Added --memberlist.address, --memberlist.advertise-address and --memberlist.members flags to configure cluster registry.
  • Added a raft cluster snapshot mode option --raft.snapshot-recovery-type which defaults to new checkpoint mode.
  • Added a flag for configuring shared table cache size --storage.table-cache-size which defaults to 1024.
  • Added support for gzip, and snappy compression for API calls.

Improvements

  • Compressed Log replication messages to lower API bandwidth usage.
  • Bump to Go 1.20.
  • Pebble instances table cache is now shared.
  • Single key lookups now utilise bloom filters optimizing the "key not present" case.
  • Added compaction metrics of pebble instances.

Bugfixes

  • Removing kafka package leftovers.
  • Improving logreader tests.
  • Fixed raft header returned from non-leader cluster nodes.
  • Fixed value of Deleted field of the DeleteRange response.
  • Fixed potential DOS by limiting the storage iterator with max RPC size.

v0.1.0

Initial Armada v0.1.0 Release

This release was made possible thanks to the contributions of @jizi @martinvizvary @jojand @mrackoa and @juldou :rocket::tada:

Special shout-out to @coufalja for his efforts!

← PreviousArchitectureNext →Contributing