Back to Blog

Speech at PGDay Israel 2026：Durability, Consistency, and Performance

NeilXP
Speech at PGDay Israel 2026：Durability, Consistency, and Performance
1
From open source to production-ready
In a traditional PostgreSQL cluster, every read replica keeps its own full copy of the data and replays the entire WAL stream, so scaling reads and failing over both pay that cost. Alibaba's PolarDB separates compute from storage instead: one writer and many readers all point at the same shared storage, so a new read replica needs no data copy and a failover needs no data resync.
At PGDay Israel 2026, Alexey Kopytov presented how we took PolarDB from a promising open-source project to production-ready Awide Polar.
📼 DataMax WAL log-store node
🧭 LSN-gated read-your-writes
⚡ Zero-copy pfsd I/O
🛡️ Disk-paxos RW lease
🔢 CSN lock-free snapshots
2
Extended Reading
01 📼
DataMax: A Dedicated WAL Log-Store Node
A node that keeps a durable, near-synchronous copy of the WAL and cascades it onward — without ever replaying a record.
Read article →
02 🧭
Read-Your-Writes at the Proxy
Replication lag breaks read-your-writes; an LSN-based wait makes a replica answer only after it has replayed far enough.
Read article →
03 ⚡
Zero-Copy I/O for the Storage Daemon
Removing the last copy from the WAL path, so the shared storage daemon writes the caller's buffer in place.
Read article →
04 🛡️
PolarFS: Disk-Paxos RW Leases
A lock written to the shared disk itself, renewed every second, backed by a watchdog that fences split-brain.
Read article →
05 🔢
Commit Sequence Number (CSN) MVCC
Lock-free snapshot acquisition: a few atomic counters replace the ProcArray walk under ProcArrayLock.
Read article →
3
Threads through the set
📼Durability is a log problem first: store the WAL, then decide whether to replay it.
🧭Consistency is a wait problem: prove a reader has caught up before it answers.
⚡Performance is a copy problem: every unnecessary copy taxes the I/O path.
🛡️Shared storage is a fencing problem: exactly one writer, enforced or terminated.
🔢Snapshots are a lock problem: read a few counters instead of walking a lock.
🎯Correctness must never be compromised for throughput—anywhere in the stack.