PostgreSQL

Change Data Capture (CDC)

Materialize supports PostgreSQL as a real-time data source. The PostgreSQL source uses PostgreSQL’s replication protocol to continually ingest changes resulting from CRUD operations in the upstream database. The native support for PostgreSQL Change Data Capture (CDC) in Materialize gives you the following benefits:

• No additional infrastructure: Ingest PostgreSQL change data into Materialize in real-time with no architectural changes or additional operational overhead. In particular, you do not need to deploy Kafka and Debezium for PostgreSQL CDC.

• Transactional consistency: The PostgreSQL source ensures that transactions in the upstream PostgreSQL database are respected downstream. Materialize will never show partial results based on partially replicated transactions.

• Incrementally updated materialized views: Materialized views in PostgreSQL are computationally expensive and require manual refreshes. You can use Materialize as a read-replica to build views on top of your PostgreSQL data that are efficiently maintained and always up-to-date.

When a source is created, Materialize parallelizes the initial snapshot across the cluster’s workers and, on PostgreSQL 14 and later, splits each table’s read across workers. See Snapshot parallelism.

Supported versions and services

The PostgreSQL source requires PostgreSQL 11+ and is compatible with most common PostgreSQL hosted services.

Integration guides

To help you get started, the following integration guides are available:

• AlloyDB for PostgreSQL
• Amazon Aurora for PostgreSQL
• Amazon RDS for PostgreSQL
• Azure DB for PostgreSQL
• Google Cloud SQL for PostgreSQL
• Neon
• Self-hosted PostgreSQL

Supported data types

Supported types

Materialize natively supports the following PostgreSQL types (including the array type for each of the types):

• bool
• bpchar
• bytea
• char
• date
• daterange
• float4
• float8
• int2
• int2vector
• int4
• int4range
• int8
• int8range
• interval
• json
• jsonb
• numeric
• numrange
• oid
• text
• time
• timestamp
• timestamptz
• tsrange
• tstzrange
• uuid
• varchar

Replicating tables that contain unsupported data types is possible via the TEXT COLUMNS option. The specified columns will be treated as text; i.e., will not have the expected PostgreSQL type features. For example:

• enum: When decoded as text, the implicit ordering of the original PostgreSQL enum type is not preserved; instead, Materialize will sort values as text.

• money: When decoded as text, resulting text value cannot be cast back to numeric, since PostgreSQL adds typical currency formatting to the output.

How ingestion from PostgreSQL works

Replication slots

Each source ingests the raw replication stream data for all tables in the specified publication using a single replication slot. To manage replication slots:

• For PostgreSQL 13+, set a reasonable value for max_slot_wal_keep_size to limit the amount of storage used by replication slots.

• If you stop using Materialize, or if either the Materialize instance or the PostgreSQL instance crash, delete any replication slots. You can query the mz_internal.mz_postgres_sources table to look up the name of the replication slot created for each source.

• If you delete all objects that depend on a source without also dropping the source, the upstream replication slot remains and will continue to accumulate data so that the source can resume in the future. To avoid unbounded disk space usage, make sure to use DROP SOURCE or manually delete the replication slot.

Snapshotting

The PostgreSQL source performs parallel snapshotting of tables by distributing rows among workers using ranges of CTID. Materialize uses PostgreSQL statistics to estimate the amount of data and number of rows to read. Missing or stale statistics can result in uneven work distribution, reducing snapshot performance. They can also cause incorrect snapshot progress reporting in the Console.

To avoid this situation, before creating the source in Materialize, ensure statistics are up to date by running PostgreSQL ANALYZE command.

Publication membership

PostgreSQL’s logical replication API does not provide a signal when users remove tables from publications. Because of this, Materialize relies on periodic checks to determine if a table has been removed from a publication, at which time it generates an irrevocable error, preventing any values from being read from the table.

However, it is possible to remove a table from a publication and then re-add it before Materialize notices that the table was removed. In this case, Materialize can no longer provide any consistency guarantees about the data we present from the table and, unfortunately, is wholly unaware that this occurred.

To mitigate this issue, if you need to drop and re-add a table to a publication, ensure that you remove the table/subsource from the source before re-adding it using the DROP SOURCE command.

Inherited tables

When using PostgreSQL table inheritance, PostgreSQL serves data from SELECTs as if the inheriting tables’ data is also present in the inherited table. However, both PostgreSQL’s logical replication and COPY only present data written to the tables themselves, i.e. the inheriting data is not treated as part of the inherited table.

PostgreSQL sources use logical replication and COPY to ingest table data, so inheriting tables’ data will only be ingested as part of the inheriting table, i.e. in Materialize, the data will not be returned when serving SELECTs from the inherited table.

• If using legacy syntax CREATE SOURCE ... FOR ...:

  You can mimic PostgreSQL’s SELECT behavior with inherited tables by creating a materialized view that unions data from the inherited and inheriting tables (using UNION ALL). However, if new tables inherit from the table, data from the inheriting tables will not be available in the view. You will need to add the inheriting tables via ADD SUBSOURCE and create a new view (materialized or non-) that unions the new table.

• If using new CREATE TABLE FROM SOURCE syntax:

  You can mimic PostgreSQL’s SELECT behavior with inherited tables by creating a materialized view that unions data from the inherited and inheriting tables (using UNION ALL). However, if new tables inherit from the table, data from the inheriting tables will not be available in the view. You will need to add the inheriting tables via CREATE TABLE .. FROM SOURCE and create a new view (materialized or non-) that unions the new table.

Partitioned tables

When you add a declaratively partitioned table to a publication, PostgreSQL expands it to the table’s leaf partitions; the parent table is not itself replicated. Materialize ingests one table per partition, which you can reassemble into the parent table using UNION ALL.

Materialize does not support ingesting from a publication created with publish_via_partition_root = true, and doing so can produce incorrect results.

See Ingest from partitioned tables for the supported approaches, including how to add and remove partitions over time.

Modifying an existing source

When you add a new subsource to an existing source (ALTER SOURCE ... ADD SUBSOURCE ...), Materialize starts the snapshotting process for the new subsource. During this snapshotting, the data ingestion for the existing subsources for the same source is temporarily blocked. As such, if possible, you can resize the cluster to speed up the snapshotting process and once the process finishes, resize the cluster for steady-state.

Supported schema and table changes

The following table summarizes how Materialize handles changes to an upstream table it is ingesting. See the details below the table for the remediation for each syntax.

This section describes how changes to upstream tables that Materialize ingests affect the corresponding Materialize tables.

Adding a column

When you add a new column to your upstream table, Materialize continues to ingest only the existing columns.

To incorporate the new column:

• If using the new CREATE SOURCE and CREATE TABLE FROM SOURCE syntax, create a new table from the source. See Handle upstream column addition.

• If using the legacy CREATE SOURCE ... FOR ... syntax that creates subsources, use DROP SOURCE to drop the affected subsource, and then add the table back to the source using ALTER SOURCE ... ADD SUBSOURCE. The re-added subsource includes the new column.

Dropping a column

Dropping columns that Materialize does not ingest (for example, columns added after the source was created, or columns that are excluded) is supported. As these columns were never ingested, you can drop them without issue.

If your Materialize source ingests a column, dropping that column from your upstream table puts the affected table into an error state.

• If using the new CREATE SOURCE and CREATE TABLE FROM SOURCE syntax, you can safely drop a column by first ignoring it in Materialize. See Handle upstream column drop.

• If using legacy CREATE SOURCE ... FOR ... syntax, use DROP SOURCE to drop the affected subsource, and then add the table back to the source using ALTER SOURCE ... ADD SUBSOURCE.

Changing constraints

Materialize ignores the following constraint changes: foreign key, CHECK, and EXCLUSION. As such, you can add or drop them without affecting ingestion.

Materialize also ignores NOT NULL, UNIQUE, and PRIMARY KEY constraints that are added after the Materialize table is created (that is, the table was created without them). Adding such a constraint, and later dropping it, does not affect ingestion.

Dropping a NOT NULL, UNIQUE, or PRIMARY KEY constraint that existed when the table was created puts the affected table into an error state.

If using the new CREATE SOURCE and CREATE TABLE FROM SOURCE syntax, you can safely drop such a constraint by first excluding it in Materialize. See Handle upstream constraint drop.

Changing a column’s data type

Changing an ingested column’s data type upstream puts the affected Materialize table into an error state unless the column was ingested as text via the TEXT COLUMNS option. Ingestion for that table stops, and you must drop and recreate the table in Materialize to resume ingestion.

Renaming a column

Renaming a column that Materialize ingests puts the affected table into an error state. Ingestion for that table stops, and you must drop and recreate the table in Materialize to resume ingestion.

Table-level operations

The following upstream operations put the affected table into an error state. Ingestion for that table stops, and you must drop and recreate the affected table in Materialize to resume:

• Dropping a table (DROP TABLE), or removing it from the publication (ALTER PUBLICATION ... DROP TABLE).
• Renaming a table or moving it to a different schema.
• Setting a table’s replica identity to anything other than FULL (ALTER TABLE ... REPLICA IDENTITY).
• Truncating a table (TRUNCATE). To clear a table without putting it into an error state, use an unqualified DELETE FROM t; instead.

Supported database operations

The following table summarizes how Materialize handles operational events on the upstream PostgreSQL database. See the details below the table for the error text and any required configuration.

Operations that do not require re-creating the source

Materialize tracks a log sequence number (LSN) as it consumes the upstream write-ahead log (WAL), and the source’s replication slot retains the WAL that Materialize has not yet consumed. Because the slot outlives the connection, routine operational events do not lose data: after a transient interruption the source stalls, then resumes from its committed LSN and catches up automatically. No action is required for the following operations:

• Restarting or patching PostgreSQL (including OS-level restarts).
• Restarting Materialize. The source resumes from its committed LSN and does not re-snapshot already-ingested data.
• Transient network interruptions between Materialize and PostgreSQL. These surface as connection closed.
• Resizing the cluster that hosts the source, or changing its replication factor. Briefly, the source may report replication slot ... is active while the upstream releases the slot from the previous connection.
• The upstream database running out of disk space, once space is freed.

Operations that require re-creating the source

A smaller set of events breaks LSN continuity or destroys the replication slot. When this happens, Materialize cannot guarantee a correct, gap-free view of your data. Most of these put the entire source into an error or permanently stalled state. One, restoring from a volume or disk snapshot, cannot be detected at all, so the source keeps running on diverged data. Every event in this section requires re-creating the source. Upstream changes to an individual table’s schema are handled separately, and do not error the entire source.

In each case below, the remediation is to drop and re-create the source:

DROP SOURCE mz_source CASCADE;

CREATE SOURCE mz_source
  FROM POSTGRES CONNECTION pg_connection (PUBLICATION 'mz_source');

-- Re-create the tables you were ingesting.
CREATE TABLE table_1 FROM SOURCE mz_source (REFERENCE public.table_1);

If you are using the legacy CREATE SOURCE ... FOR TABLES syntax, re-create the source with the same FOR TABLES list instead of adding tables separately.

Because a re-created source snapshots from the current state of the upstream database, any changes it missed while it was in an error state are reflected in the snapshot rather than replayed as individual updates.

Point-in-time restore

Restoring the source database from a backup, including restoring to a different server for disaster recovery, increments the PostgreSQL timeline and is detected as a discontinuity. The source fails with an error of the form:

unsupported action: database restored from point-in-time backup. Expected
timeline ID 8 but got 9

The same error covers other events that change the timeline, such as a managed failover between replicas. To see the timeline a source is pinned to, query mz_internal.mz_postgres_sources:

SELECT s.name, p.replication_slot, p.timeline_id
FROM mz_internal.mz_postgres_sources p
JOIN mz_catalog.mz_sources s ON s.id = p.id;

If your upstream fails over between replicas as part of routine maintenance, see High-availability failovers.

Restoring from a volume or disk snapshot

Restoring the upstream data directory from a crash-consistent volume or disk snapshot rolls the database back, but preserves the timeline ID and the replication slot. Materialize cannot detect this kind of restore. The source keeps running without an error, but its contents diverge from upstream. This can surface later as incorrect results, or as negative-accumulation errors in queries such as Non-positive multiplicity.

Promotion of a physical replica

When a source reads from a physical standby (read replica) rather than the primary, promoting that standby to a primary fails the source with:

unsupported action: upstream physical replica status changed (e.g. a physical
replica was promoted to a primary). Expected pg_is_in_recovery()=true but got
false

Materialize detects the promotion while the replication stream is live, without waiting for a restart. Re-create the source against the promoted node.

Replication slot invalidated

PostgreSQL invalidates a replication slot once the WAL it holds exceeds max_slot_wal_keep_size. This protects the upstream from running out of disk, at the cost of ending replication. The source fails with:

replication slot has been invalidated because it exceeded the maximum reserved
size

To avoid this, size the source cluster so that it keeps up with the upstream write rate, and set max_slot_wal_keep_size high enough to cover your longest expected outage. Some hosted PostgreSQL services set this value for you and do not allow it to be raised.

Replication slot dropped or rewound

If the slot Materialize is using is dropped upstream, or the upstream is rebuilt from a base backup (which does not carry replication slots), a new slot starts at the current LSN, past the point the source needs to resume from. The source stalls with:

slot overcompacted. Requested LSN ... but only LSNs >= ... are available

For diagnosis steps, see Slot overcompacted. PostgreSQL refuses to drop a slot that is in use, so this generally happens only while the source is paused or disconnected.

Not every rewind is caught this way. A rewind that leaves the slot able to serve the LSN the source asks for, such as restoring from a volume or disk snapshot, raises no error at all.

Dropping the publication

Running DROP PUBLICATION upstream stalls the source, and all of its tables, with:

publication "mz_source" does not exist

Re-create the publication upstream, then re-create the source.

Major version upgrades

A PostgreSQL major version upgrade rewrites the on-disk format and does not preserve the replication slot, so there is no in-place recovery. To upgrade without a gap in your downstream views, run a second source against the upgraded instance in parallel and cut over once it has hydrated. See Upgrade the major version of your PostgreSQL source.

High-availability failovers

Some managed PostgreSQL services increment the timeline during routine high-availability operations, such as maintenance, a machine-tier change, or an automatic failover between replicas. Materialize cannot distinguish these from a genuine restore, so by default they fail the source with the Expected timeline ID error.

On self-managed Materialize, where the upstream service guarantees that a failover is a contiguous fork of the WAL with no data loss, you can disable timeline validation with the pg_source_validate_timeline system parameter:

ALTER SYSTEM SET pg_source_validate_timeline = false;

This parameter is not available on Materialize Cloud. There, a high-availability failover that changes the timeline requires re-creating the source.