Postgres_cluster

Various experiments with PostgreSQL clustering perfomed at PostgresPro.

This is mirror of postgres repo with several changes to the core and few extra extensions.

Core changes:

Transaction manager interface (eXtensible Transaction Manager, xtm). Generic interface to plug distributed transaction engines. More info at [[https://wiki.postgresql.org/wiki/DTM]] and [[http://www.postgresql.org/message-id/flat/F2766B97-555D-424F-B29F-E0CA0F6D1D74@postgrespro.ru]].
Distributed deadlock detection API.
Logical decoding of two-phase transactions.

New extensions:

pg_tsdtm. Coordinator-less transaction management by tracking commit timestamps.
multimaster. Synchronous multi-master replication based on logical_decoding and pg_dtm.

Changed extension:

postgres_fdw. Added support of pg_tsdtm.

Installing multimaster

Build and install postgres from this repo on all machines in cluster.
Install contrib/raftable and contrib/mmts extensions.
Right now we need clean postgres installation to spin up multimaster cluster.
Create required database inside postgres before enabling multimaster extension.
We are requiring following postgres configuration:
1. 'max_prepared_transactions' > 0 -- in multimaster all writing transaction along with ddl are wrapped as two-phase transaction, so this number will limit maximum number of writing transactions in this cluster node.
2. 'synchronous_commit - off' -- right now we do not support async commit. (one can enable it, but that will not bring desired effect)
3. 'wal_level = logical' -- multimaster built on top of logical replication so this is mandatory.
4. 'max_wal_senders' -- this should be at least number of nodes - 1
5. 'max_replication_slots' -- this should be at least number of nodes - 1
6. 'max_worker_processes' -- at least 2*N + P - 1, where N is number of nodes in cluster, P size of pool of workers(see below) (1 raftable, n-1 receiver, n-1 sender, mtm-sender, mtm-receiver, + number of pool worker).
7. 'default_transaction_isolation = 'repeatable read'' -- multimaster isn't supporting default read commited level.
Also multimaster have following configuration parameters:
1. 'multimaster.conn_strings' -- connstrings for all nodes in cluster, separated by comma.
2. 'multimaster.node_id' -- id of current node, number starting from one.
3. 'multimaster.workers' -- number of workers that can apply transactions from neighbouring nodes.
4. 'multimaster.use_raftable = true' -- just set this to true. Deprecated.
5. 'multimaster.queue_size = 52857600' -- queue size for applying transactions from neighbouring nodes.
6. 'multimaster.ignore_tables_without_pk = 1' -- do not replicate tables without primary key
7. 'multimaster.heartbeat_send_timeout = 250' -- heartbeat period (ms).
8. 'multimaster.heartbeat_recv_timeout = 1000' -- disconnect node if we miss heartbeats all that time (ms).
9. 'multimaster.twopc_min_timeout = 40000' -- rollback stalled transaction after this period (ms).
10. 'raftable.id' -- id of current node, number starting from one.
11. 'raftable.peers' -- id of current node, number starting from one.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!