Distributed Postgres
1 like393 views
This document discusses distributed Postgres including multi-master replication, distributed transactions, and high availability/auto failover. It explores existing implementations like Postgres-XC and proposes a transaction manager API and time-stamp based approach to enable distributed transactions without a central bottleneck. The document also outlines a multimaster implementation built on logical replication, a transaction replay pool, and Raft-based storage for failure handling and distributed deadlocks. Performance is approximately half of standalone Postgres with the same read speeds and capabilities for node recovery and network partition handling.
![More nodes, higher probability of failure in system.
Possible problems with nodes:
Node stopped (and will not be back);
Node was down small amount of time (and we should bring it
back to operation);
Network partitions (avoid split-brain).
If we want to survive network partitions than we can have not more
than [N/2] - 1 failures.
HA/autofailover
13](https://image.slidesharecdn.com/distributedpostgres-160518144919/85/Distributed-Postgres-13-320.jpg)
Distributed Postgres
- 1. Distributed postgres. XL, XTM, MultiMaster Stas Kelvich
- 2. Started about a year ago. Konstantin Knizhnik, Constantin Pan, Stas Kelvich Cluster group in PgPro 2
- 3. Started to playing with Postgres-XC. 2ndQuadrant also had project (finished now) to port XC to 9.5. Fork is painful; How can we bring functionality of XC in core? Cluster group in PgPro 3
- 4. Distributed transactions - nothing in-core; Distributed planner - fdw, pg_shard, greenplum planner (?); HA/Autofailover - can be built on top of logical decoding. Distributed postgres 4
- 5. Achieve proper isolation between tx for multi-node transactions. Now in postgres on write tx start: Aquire XID; Get list of running tx’s; Use that info in visibility checks. Distributed transactions 5
- 9. Aquire XID centrally (DTMd, arbiter); No local tx possible; DTMd is a bottleneck. XTM implementations GTM or snapshot sharing 9
- 10. Paper from SAP HANA team; Central daemon is needed, but only for multi-node tx; Snapshots -> Commit Sequence Number; DTMd is still a bottleneck. XTM implementations Incremental SI 10
- 11. XID/CSN are gathered from all nodes that participates in tx; No central service; local tx; possible to reduce communication by using time (Spanner, CockroachDB). XTM implementations Clock-SI or tsDTM 11
- 13. More nodes, higher probability of failure in system. Possible problems with nodes: Node stopped (and will not be back); Node was down small amount of time (and we should bring it back to operation); Network partitions (avoid split-brain). If we want to survive network partitions than we can have not more than [N/2] - 1 failures. HA/autofailover 13
- 14. Possible usage of such system: Multimaster replication; Tables with metainformation in sharded databases; Sharding with redundancy. HA/autofailover 14
- 15. By Multimaster we mean strongly coupled one, that acts as a single database. With proper isolation and no merge conflicts. Ways to build: Global order to XLOG (Postgres-R, MySQL Galera); Wrap each tx as distributed – allows parallelism while applying tx. Multimaster 15
- 16. Our implementation: Built on top of pg_logical; Make use of tsDTM; Pool of workers for tx replay; Raft-based storage for dealing with failures and distributed deadlock detection. Multimaster 16
- 17. Our implementation: Approximately half of a speed of standalone postgres; Same speed for reads; Deals with nodes autorecovery; Deals with network partitions (debugging right now). Can work as an extension (if community accept XTM API in core). Multimaster 17