Skip to main content

Amazon RDS for PostgreSQL

This connector uses change data capture (CDC) to continuously capture updates in a PostgreSQL database into one or more Flow collections.

It is available for use in the Flow web application. For local development or open-source workflows, ghcr.io/estuary/source-postgres:dev provides the latest version of the connector as a Docker image. You can also follow the link in your browser to see past image versions.

Supported versions and platforms​

This connector supports PostgreSQL versions 10.0 and later on major cloud platforms.

Prerequisites​

You'll need a PostgreSQL database setup with the following:

  • Logical replication enabled — wal_level=logical
  • User role with REPLICATION attribute
  • A replication slot. This represents a “cursor” into the PostgreSQL write-ahead log from which change events can be read.
    • Optional; if none exist, one will be created by the connector.
    • If you wish to run multiple captures from the same database, each must have its own slot. You can create these slots yourself, or by specifying a name other than the default in the advanced configuration.
  • A publication. This represents the set of tables for which change events will be reported.
    • In more restricted setups, this must be created manually, but can be created automatically if the connector has suitable permissions.
  • A watermarks table. The watermarks table is a small “scratch space” to which the connector occasionally writes a small amount of data to ensure accuracy when backfilling preexisting table contents.
    • In more restricted setups, this must be created manually, but can be created automatically if the connector has suitable permissions.

Setup​

  1. Allow connections between the database and Estuary Flow. There are two ways to do this: by granting direct access to Flow's IP or by creating an SSH tunnel.

    1. To allow direct access:

    2. To allow secure connections via SSH tunneling:

  2. Enable logical replication on your RDS PostgreSQL instance.

    1. Create a parameter group. Create a unique name and description and set the following properties:

      • Family: postgres13
      • Type: DB Parameter group

    2. Modify the new parameter group and set rds.logical_replication=1.

    3. Associate the parameter group with the database.

    4. Reboot the database to allow the new parameter group to take effect.

  3. In the PostgreSQL client, connect to your instance and run the following commands to create a new user for the capture with appropriate permissions, and set up the watermarks table and publication.

CREATE USER flow_capture WITH PASSWORD 'secret';
GRANT rds_replication TO flow_capture;
GRANT SELECT ON ALL TABLES IN SCHEMA public TO flow_capture;
ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT SELECT ON TABLES TO flow_capture;
CREATE TABLE IF NOT EXISTS public.flow_watermarks (slot TEXT PRIMARY KEY, watermark TEXT);
GRANT ALL PRIVILEGES ON TABLE public.flow_watermarks TO flow_capture;
CREATE PUBLICATION flow_publication;
ALTER PUBLICATION flow_publication SET (publish_via_partition_root = true);
ALTER PUBLICATION flow_publication ADD TABLE public.flow_watermarks, <other_tables>;

where <other_tables> lists all tables that will be captured from. The publish_via_partition_root setting is recommended (because most users will want changes to a partitioned table to be captured under the name of the root table) but is not required.

  1. In the RDS console, note the instance's Endpoint and Port. You'll need these for the address property when you configure the connector.

Backfills and performance considerations​

When the a PostgreSQL capture is initiated, by default, the connector first backfills, or captures the targeted tables in their current state. It then transitions to capturing change events on an ongoing basis.

This is desirable in most cases, as in ensures that a complete view of your tables is captured into Flow. However, you may find it appropriate to skip the backfill, especially for extremely large tables.

In this case, you may turn of backfilling on a per-table basis. See properties for details.

Configuration​

You configure connectors either in the Flow web app, or by directly editing the catalog specification file. See connectors to learn more about using connectors. The values and specification sample below provide configuration details specific to the PostgreSQL source connector.

Properties​

Endpoint​

PropertyTitleDescriptionTypeRequired/Default
/addressAddressThe host or host:port at which the database can be reached.stringRequired
/databaseDatabaseLogical database name to capture from.stringRequired, "postgres"
/userUserThe database user to authenticate as.stringRequired, "flow_capture"
/passwordPasswordPassword for the specified database user.stringRequired
/advancedAdvanced OptionsOptions for advanced users. You should not typically need to modify these.object
/advanced/backfill_chunk_sizeBackfill Chunk SizeThe number of rows which should be fetched from the database in a single backfill query.integer4096
/advanced/publicationNamePublication NameThe name of the PostgreSQL publication to replicate from.string"flow_publication"
/advanced/skip_backfillsSkip BackfillsA comma-separated list of fully-qualified table names which should not be backfilled.string
/advanced/slotNameSlot NameThe name of the PostgreSQL replication slot to replicate from.string"flow_slot"
/advanced/watermarksTableWatermarks TableThe name of the table used for watermark writes during backfills. Must be fully-qualified in '<schema>.<table>' form.string"public.flow_watermarks"
/advanced/sslmodeSSL ModeOverrides SSL connection behavior by setting the 'sslmode' parameter.string

Bindings​

PropertyTitleDescriptionTypeRequired/Default
/namespaceNamespaceThe namespace/schema of the table.stringRequired
/streamStreamTable name.stringRequired
/syncModeSync modeConnection method. Always set to incremental.stringRequired

SSL Mode​

Certain managed PostgreSQL implementations may require you to explicitly set the SSL Mode to connect with Flow. One example is Neon, which requires the setting verify-full. Check your managed PostgreSQL's documentation for details if you encounter errors related to the SSL mode configuration.

Sample​

A minimal capture definition will look like the following:

captures:
  ${PREFIX}/${CAPTURE_NAME}:
    endpoint:
      connector:
        image: "ghcr.io/estuary/source-postgres:dev"
        config:
          address: "localhost:5432"
          database: "postgres"
          user: "flow_capture"
          password: "secret"
    bindings:
      - resource:
          stream: ${TABLE_NAME}
          namespace: ${TABLE_NAMESPACE}
          syncMode: incremental
        target: ${PREFIX}/${COLLECTION_NAME}

Your capture definition will likely be more complex, with additional bindings for each table in the source database.

Learn more about capture definitions.

TOASTed values​

PostgreSQL has a hard page size limit, usually 8 KB, for performance reasons. If your tables contain values that exceed the limit, those values can't be stored directly. PostgreSQL uses TOAST (The Oversized-Attribute Storage Technique) to store them separately.

TOASTed values can sometimes present a challenge for systems that rely on the PostgreSQL write-ahead log (WAL), like this connector. If a change event occurs on a row that contains a TOASTed value, but the TOASTed value itself is unchanged, it is omitted from the WAL. As a result, the connector emits a row update with the a value omitted, which might cause unexpected results in downstream catalog tasks if adjustments are not made.

The PostgreSQL connector handles TOASTed values for you when you follow the standard discovery workflow or use the Flow UI to create your capture. It uses merge reductions to fill in the previous known TOASTed value in cases when that value is omitted from a row update.

However, due to the event-driven nature of certain tasks in Flow, it's still possible to see unexpected results in your data flow, specifically:

  • When you materialize the captured data to another system using a connector that requires delta updates
  • When you perform a derivation that uses TOASTed values

Troubleshooting​

If you encounter an issue that you suspect is due to TOASTed values, try the following:

Publications​

It is recommended that the publication used by the capture only contain the tables that will be captured. In some cases it may be desirable to create this publication for all tables in the database instead of specific tables, for example using:

CREATE PUBLICATION flow_publication FOR ALL TABLES WITH (publish_via_partition_root = true);

Caution must be used if creating the publication in this way as all existing tables (even those not part of the capture) will be included in it, and if any of them do not have a primary key they will no longer be able to process updates or deletes.