Snowflake destination

Snowflake is a managed data platform. Supabase Pipelines writes an append-only change history for each replicated Postgres table to Snowflake.

Prepare Snowflake resources#

Create a dedicated Snowflake database, schema, role, and service user for Pipelines. Keep the schema otherwise empty to avoid ownership conflicts. Use unquoted identifiers for the service user and role. Pipelines converts the account and user names to uppercase during authentication.

Run the following as a Snowflake administrator. Change the example names as needed:

1
create role if not exists PIPELINES_ROLE;
2
3
create user if not exists PIPELINES_USER
4
  type = service;
5
6
grant role PIPELINES_ROLE to user PIPELINES_USER;
7
alter user PIPELINES_USER set default_role = PIPELINES_ROLE;
8
9
create database if not exists PIPELINES_DB;
10
create schema if not exists PIPELINES_DB.REPLICATED;
11
12
grant usage on database PIPELINES_DB to role PIPELINES_ROLE;
13
grant usage on schema PIPELINES_DB.REPLICATED to role PIPELINES_ROLE;
14
grant create table on schema PIPELINES_DB.REPLICATED to role PIPELINES_ROLE;
15
grant create pipe on schema PIPELINES_DB.REPLICATED to role PIPELINES_ROLE;

The pipeline role must own destination tables so it can alter, truncate, or drop them. Don't pre-create destination tables under another role.

The role also needs CREATE PIPE so Snowflake can create each table's managed default Snowpipe Streaming pipe, named <TABLE>-STREAMING, when Pipelines opens a channel. Pipelines does not need a virtual warehouse, stage, or manually created pipe.

Use a separate role and warehouse for downstream queries and transformations. The pipeline service role does not need query or transformation privileges.

Keep the SQL and streaming roles aligned#

Pipelines uses two Snowflake interfaces:

• SQL requests use the optional Role configured in the Dashboard. When Role is empty, they use the user's default role.
• Snowpipe Streaming uses the user's DEFAULT_ROLE. It does not use the optional Role setting.

Use one dedicated role for both interfaces. Set it as the service user's DEFAULT_ROLE. Leave Role empty in the Dashboard or set it to the same role. If the roles differ, SQL validation and table creation can succeed while streaming fails.

Generate a key pair#

Pipelines authenticates with an RSA key pair. Snowflake requires a key of at least 2048 bits and recommends PKCS #8. Generate an unencrypted private key:

1
openssl genrsa 2048 | openssl pkcs8 -topk8 -inform PEM -out rsa_key.p8 -nocrypt

To use a passphrase, generate an encrypted PKCS #8 private key:

1
openssl genrsa 2048 | openssl pkcs8 -topk8 -v2 des3 -inform PEM -out rsa_key.p8

Derive the public key:

1
openssl rsa -in rsa_key.p8 -pubout -out rsa_key.pub

Register only the public-key body with the service user. Omit the BEGIN PUBLIC KEY and END PUBLIC KEY lines:

1
alter user PIPELINES_USER set rsa_public_key = '<public-key-body>';

Keep rsa_key.p8 and its passphrase secret. Don't commit them, paste them into logs, or send them to support. The Dashboard accepts unencrypted PKCS #1 or PKCS #8 keys, and encrypted PKCS #8 keys with a passphrase. It does not support encrypted PKCS #1 keys.

See Snowflake key-pair authentication to verify the public-key fingerprint and rotate keys with RSA_PUBLIC_KEY_2.

Find the account identifier#

Run this query in Snowflake:

1
select current_organization_name() || '-' || current_account_name();

Enter the result as Account ID, for example MYORG-MYACCOUNT. Do not enter a full URL or dotted locator-and-region hostname. Account IDs can contain up to 63 characters. Legacy one-part account locators are also accepted. See Snowflake account identifiers for details.

Configure Snowflake as a destination#

1. Navigate to the Database > Replication section of the Dashboard.
2. Click Add destination.
3. Select Snowflake. If it isn't available, request Early Access.
4. Select a Postgres publication and enter a destination name.
5. Enter the Snowflake settings:
   • Account ID: The organization and account identifier, such as MYORG-MYACCOUNT.
   • User: The dedicated unquoted service user, such as PIPELINES_USER.
   • Database: The destination database, such as PIPELINES_DB.
   • Schema: The dedicated destination schema, such as REPLICATED.
   • Role: Leave empty to use the user's DEFAULT_ROLE. If set, enter the same role.
   • Private key: The complete PEM-encoded private key, including its begin and end lines.
   • Private key passphrase: Required only for an encrypted PKCS #8 key.
6. Review the source table requirements and click Create and start pipeline.

Enter the database and schema identifiers exactly as stored in Snowflake. Unquoted identifiers are stored in uppercase. Managed Pipelines run in AWS eu-central-1 (Frankfurt). When possible, use a Snowflake account near Frankfurt.

How it works#

Pipelines uses Snowflake's SQL REST API to validate the database and schema, create, evolve, and reset destination tables, and apply source TRUNCATE operations. It sends initial and ongoing row data through Snowpipe Streaming. These operations do not use a virtual warehouse.

Validation checks authentication, database and schema visibility, and that QUOTED_IDENTIFIERS_IGNORE_CASE is FALSE. It does not verify that the role can create or own tables, create pipes, or write through Snowpipe Streaming.

Destination table names#

Pipelines maps each Postgres schema and table pair to one Snowflake table name. It doubles existing underscores, joins the names with one underscore, and uppercases the result:

Postgres schema and table names cannot start or end with _ or contain " or ;. Names that differ only in case map to the same Snowflake name. Use lowercase Postgres names to avoid collisions. Source column names are preserved as quoted identifiers, except for the reserved metadata names below.

Append-only change history#

Each destination table contains the replicated source columns plus two VARCHAR NOT NULL metadata columns:

The metadata names are reserved and can't be used by source columns. Initial-sync rows use insert and the shared sequence number 0000000000000000/0000000000000000.

Snowflake tables are an event history, not a current-state replica:

• An insert appends the new row.
• An update appends the complete new row. It does not append a before image.
• A delete appends the complete old row for REPLICA IDENTITY FULL. For a primary-key or USING INDEX identity, it appends only the identity columns and sets all other source columns to NULL.
• A source TRUNCATE truncates the Snowflake table, resets its streaming state, and does not append a truncate event.

To derive current state, group by a stable source identity and select the row with the latest _cdc_sequence_number. Exclude identities whose latest operation is delete. The sequence number is used for ordering and checkpointing. It is not a globally unique event ID. Pipelines provides at-least-once delivery, so consumers must tolerate duplicates. Snowpipe committed offsets suppress routine replay but do not change this guarantee.

Resetting a table drops and recreates its Snowflake table and managed streaming state. This erases its history. Removing a table from the Postgres publication stops new changes after the pipeline restarts. The existing Snowflake table remains.

Source table requirements#

Required REPLICA IDENTITY depends on the operations enabled in the Postgres publication:

Set full replica identity before publishing updates:

1
alter table public.your_table replica identity full;

REPLICA IDENTITY FULL increases WAL volume, but lets Pipelines construct complete new rows when Postgres omits unchanged out-of-line TOAST values. The setting applies only to new WAL records. If retained WAL already contains an incompatible update, reset the affected table after changing the setting.

Type mapping#

Pipelines creates Snowflake columns with these mappings:

Pipelines uses VARCHAR for character and text types, numeric, time with time zone, interval, uuid, bytea, bit strings, and custom or unknown types. bytea values are lowercase hexadecimal strings. Pipelines stores these values in serialized form, not as native Snowflake types.

Additional limits apply:

• Multi-dimensional arrays aren't supported. Non-default lower bounds on one-dimensional arrays aren't preserved.
• Non-finite floating-point and numeric values are rejected.
• An uncompressed serialized row larger than 2 MiB is rejected.
• Source primary-key, unique, check, length, precision, and nullability constraints aren't copied. Only the two CDC metadata columns are NOT NULL.

Schema change support#

Snowflake schema change support is limited during Early Access.

Supported changes:

• Add a column.
• Rename a column.
• Drop a column.

Unsupported or limited changes:

• Changing a column type isn't supported.
• Source table and schema renames aren't supported.
• Changes to nullability or existing column defaults are ignored.
• Initial table creation can copy compatible literal defaults. Added columns can copy string, numeric, or boolean literal defaults. Other defaults are omitted.

Snowflake DDL changes existing history. Adding a column with a default can populate older rows. Renaming a column changes the historical schema. Dropping a column removes it from old events. Snowflake DDL is not transactional, so an interrupted multi-column change can leave a partially applied schema. Do not alter managed destination objects manually. If the pipeline remains failed after a restart, contact support.

Troubleshooting#

Use pipeline monitoring and replication logs to inspect table state, lag, and errors.

Additional resources#

• Snowflake key-pair authentication
• Snowflake account identifiers
• Snowpipe Streaming default pipe
• Snowpipe Streaming operations and privileges