pd-server v2.1.0

• Optimize availability

  • Introduce the version control mechanism and support rolling update of the cluster compatibly
  • Enable Raft PreVote among PD nodes to avoid leader reelection when network recovers after network isolation
  • Enable raft learner by default to lower the risk of unavailable data caused by machine failure during scheduling
  • TSO allocation is no longer affected by the system clock going backwards
  • Support the Region merge feature to reduce the overhead brought by metadata

• Optimize the scheduler

  • Optimize the processing of Down Store to speed up making up replicas
  • Optimize the hotspot scheduler to improve its adaptability when traffic statistics information jitters
  • Optimize the start of Coordinator to reduce the unnecessary scheduling caused by restarting PD
  • Optimize the issue that Balance Scheduler schedules small Regions frequently
  • Optimize Region merge to consider the number of rows within the Region
  • Add more commands to control the scheduling policy
  • Improve PD simulator to simulate the scheduling scenarios

• API and operation tools

  • Add the GetPrevRegion interface to support the TiDB reverse scan feature
  • Add the BatchSplitRegion interface to speed up TiKV Region splitting
  • Add the GCSafePoint interface to support distributed GC in TiDB
  • Add the GetAllStores interface, to support distributed GC in TiDB
  • pd-ctl supports:
    • using statistics for Region split
    • calling jq to format the JSON output
    • checking the Region information of the specified store
    • checking topN Region list sorted by versions
    • checking topN Region list sorted by size
    • more precise TSO encoding
  • pd-recover doesn't need to provide the max-replica parameter

• Metrics

  • Add related metrics for Filter
  • Add metrics about etcd Raft state machine

• Performance

  • Optimize the performance of Region heartbeat to reduce the memory overhead brought by heartbeats
  • Optimize the Region tree performance
  • Optimize the performance of computing hotspot statistics

pd-server v2.0.9

• Fix the issue that the PD server gets stuck caused by etcd startup failure #1267
• Fix the issues related to pd-ctl reading the Region key #1298 #1299 #1308
• Fix the issue that the regions/check API returns the wrong result #1311
• Fix the issue that PD cannot restart join after a PD join failure #1279

pd-server v2.1.0-rc.5

• Fix the issues related to pd-ctl reading the Region key #1298 #1299 #1308
• Fix the issue that the regions/check API returns the wrong result #1311
• Fix the issue that PD cannot restart join after a PD join failure 1279
• Fix the issue that watch leader might lose events in some cases 1317

pd-server v2.1.0-rc.4

• Fix the issue that the tombstone TiKV is not removed from Grafana #1261
• Fix the data race issue when grpc-go configures the status #1265
• Fix the issue that the PD server gets stuck caused by etcd startup failure #1267
• Fix the issue that data race might occur during leader switching #1273
• Fix the issue that extra warning logs might be output when TiKV becomes tombstone #1280

pd-server v2.1.0-rc.3

New feature

• Add the API to get the Region list by size in reverse order #1254

Improvement

• Return more detailed information in the Region API #1252

Bug fix

• Fix the issue that adjacent-region-scheduler might lead to a crash after PD switches the leader #1250

pd-server v2.1.0-rc.2

Features

• Support the GetAllStores interface
• Add the statistics of scheduling estimation in Simulator

Improvements

• Optimize the handling process of down stores to make up replicas as soon as possible
• Optimize the start of Coordinator to reduce the unnecessary scheduling caused by restarting PD
• Optimize the memory usage to reduce the overhead caused by heartbeats
• Optimize error handling and improve the log information
• Support querying the Region information of a specific store in pd-ctl
• Support querying the topN Region information based on version
• Support more accurate TSO decoding in pd-ctl

Bug fix

• Fix the issue that pd-ctl uses the hot store command to exit wrongly

pd-server v2.1.0-rc.1

Features

• Introduce the version control mechanism and support rolling update of the cluster with compatibility
• Enable the region merge feature
• Support the GetPrevRegion interface
• Support splitting Regions in batch
• Support storing the GC safepoint

Improvements

• Optimize the issue that TSO allocation is affected by the system clock going backwards
• Optimize the performance of handling Region hearbeats
• Optimize the Region tree performance
• Optimize the performance of computing hotspot statistics
• Optimize returning the error code of API interface
• Add options of controlling scheduling strategies
• Prohibit using special characters in label
• Improve the scheduling simulator
• Support splitting Regions using statistics in pd-ctl
• Support formatting JSON output by calling jq in pd-ctl
• Add metrics about etcd Raft state machine

Bug fixes

• Fix the issue that the namespace is not reloaded after switching Leader
• Fix the issue that namespace scheduling exceeds the schedule limit
• Fix the issue that hotspot scheduling exceeds the schedule limit
• Fix the issue that wrong logs are output when the PD client closes
• Fix the wrong statistics of Region heartbeat latency

pd-server v2.0.5

Bug Fixes

• Fix the issue that replicas migration uses up TiKV disks space in some scenarios
• Fix the crash issue caused by AdjacentRegionScheduler

pd-server v2.1.0-beta

Improvements

• Enable Raft PreVote between PD nodes to avoid leader reelection when network recovers after network isolation
• Optimize the issue that Balance Scheduler schedules small Regions frequently
• Optimize the hotspot scheduler to improve its adaptability in traffic statistics information jitters
• Skip the Regions with a large number of rows when scheduling region merge
• Enable raft learner by default to lower the risk of unavailable data caused by machine failure during scheduling
• Remove max-replica from pd-recover
• Add Filter metrics

Bug Fixes

• Fix the issue that Region information is not updated after tikv-ctl unsafe recovery
• Fix the issue that TiKV disk space is used up caused by replica migration in some scenarios

Compatibility notes

• Do not support rolling back to v2.0.x or earlier due to update of the new version storage engine
• Enable raft learner by default in the new version of PD. If the cluster is upgraded from 1.x to 2.1, the machine should be stopped before upgrade or a rolling update should be first applied to TiKV and then PD

pd-server v2.0.4

• Improve the behavior of the unset scheduling argument max-pending-peer-count by changing it to no limit for the maximum number of PendingPeers