RawKV CDC Prototype Development Task

[pingyu]

Development Task

On branch https://github.com/pingyu/tikv/tree/feat-rawkv-cdc-prototype, based on tag v5.0.4-20211201

Objectives

1. Validate correctness of proposal
2. Gain less than 10% performance regression by a coarse implementation

Sub Tasks

Totally: 2 weeks
TiKV: 1 weeks, BR/TiCDC/Observation: 0.5 week, Testing: 0.5 week

• TiKV (1 week)
  • causal timestamp (0.2 week)
    • RPC with PD
    • next ts
    • causally advance
    • cache management (a HLC likeness)
  • raw_put (0.5 week)
    • service tier
    • value encoding
    • raftstore tier
      • cb, raft message
      • Region-Max-TS
      • follower capture ts from raft
        • msg append
        • snapshot
      • ts advance on leader transfer
      • split
      • merge
  • raw_delete
  • read: deleted flag
  • cdc (0.3 week)
    • raw CmdObserver
    • incremental scan
    • RPC with TiCDC
  • resolved_ts
    • advance
    • improvement by no locking
  • Performance metrics
    • pre-propose coprocessor
• BR (0.1 week)
  • functional verification
  • command line arguments
  • deleted flag
• TiCDC (0.4 week)
  • raw changefeed
  • raw task
  • upstream RPC from TikV
  • downstream RPC to TiKV
  • checkpoint
  • RPO capture, calculation & report
  • scheduler by key span
• Prom & Grafana
  • RPO monitoring
• PD (no change)
• Test (0.5 week)
  • Correctness validation
    • Workload: N threads x M keys, PUT (k{m}, v), (k{m}, v+1). Verify values of all N x M keys are v+1. N=400, M=1000
    • Case 1: Workload on a main (1 PD + 1 KV + 1 TiCDC) + recovery (1 PD + 1 KV) clusters.
    • Case 2: The deployment as Case 1, but as TiCDC startup BETWEEN & AFTER workload.
    • Case 3: Workload on a main (1 PD + 3 KV + 1 TiCDC) + recovery (1 PD + 1 KV) clusters, trigger the following:
      • Case 3.1: Region SPLIT.
      • Case 3.2: Region MERGE.
      • Case 3.3: Leader TRANSFER.
  • Benchmark
    • Case 1: YCSB workloada (50 read / 50 write). Observing write QPS, P99 latency, CDC latency.