Skip to content

estuary/agent-skills / skills/materialize-clickhouse-create

materialize-clickhouse-create

Verified

Create a ClickHouse materialization to stream Estuary collections into ClickHouse tables. Use when setting up ClickHouse as a destination for captured data. Use when user says "send to ClickHouse", "materialize to ClickHouse", "ClickHouse destination", or "load into ClickHouse".

Create ClickHouse Materialization

Create a ClickHouse materialization using flowctl to stream data from Estuary collections into ClickHouse tables.

Applies to: materialize-clickhouse (self-hosted and ClickHouse Cloud)

Step 0: Load Connector Documentation

Before proceeding, fetch the official connector docs for prerequisites, config reference, and setup instructions.

Load the docs page: https://docs.estuary.dev/reference/Connectors/materialization-connectors/ClickHouse/

Use WebFetch to load this page. It covers:

  • Prerequisites (user creation, database and system table permissions)
  • Full config property reference
  • SSL configuration (sslmode: disable, require, verify-full)
  • ReplacingMergeTree engine details and FINAL directive usage
  • Advanced options (delta updates, hard deletes)

Search Kapa for tribal knowledge (if the Estuary MCP is configured):

Search kapa ai knowledge sources for "materialize clickhouse common issues"

If Kapa MCP is not configured, the user can set it up: https://docs.estuary.dev/features/mcp-integration/

This skill provides the flowctl workflow and troubleshooting that docs don't cover.

Step 1: Gather Requirements

Before writing any YAML, ask the user:

  1. ClickHouse deployment? — Self-hosted or ClickHouse Cloud
  2. Address? — host[:port]. Default port is 9440 with TLS enabled (default) or 9000 with TLS disabled. The connector uses the Native protocol — not the HTTP interface on port 8123.
  3. Target database? — Name of the ClickHouse database to materialize to
  4. Username and password? — Auth type is user_password (the only supported auth method)
  5. SSL mode? — Defaults to verify-full. Options: disable, require, verify-full
  6. Non-default data plane? — Most users use the default. Ask if they need a non-default data plane.
  7. Source collections? — Which Estuary collections to materialize
  8. Hard deletes? — Off by default (soft-delete via _meta/op column). Enable to insert tombstone rows (_is_deleted = 1) that ClickHouse uses to exclude rows from FINAL queries and eventually remove from storage.
  9. Delta updates? — Off by default. When enabled per-binding, the table uses the MergeTree engine instead of ReplacingMergeTree. Every store is appended as-is with no deduplication. Use for event logs or history tables.

CRITICAL: Use the Native protocol, not HTTP. Port 8123 (HTTP interface) is not supported. Use port 9440 (TLS) or 9000 (no TLS).

CRITICAL: Query with the FINAL directive. In standard (non-delta) mode, the connector uses ReplacingMergeTree with flow_published_at as the version column. Duplicate keys and tombstones are deduplicated in background merges — queries must use SELECT ... FROM my_table FINAL to see correctly deduplicated results.

CRITICAL: ClickHouse allow_nullable_key must be enabled. The connector creates tables with ORDER BY (_meta/row_id) where _meta/row_id is Nullable(Int64). ClickHouse refuses nullable sort-key columns by default and the first CREATE TABLE fails with Code 44: Sorting key contains nullable columns. See Step 3 for the fix — ClickHouse Cloud has this set already; self-hosted servers need a config snippet.

Load pattern note (relevant if you build downstream materialized views)

The connector does not use plain INSERT INTO <table>. For each batch it stages rows into flow_temp_store_0_<table> and then issues ALTER TABLE flow_temp_store_0_<table> MOVE PARTITION ID 'all' TO TABLE <table>. MOVE PARTITION is a metadata operation — it relinks part files atomically without re-writing data.

Two consequences worth telling the user upfront:

  • Batch loads are essentially free at the ClickHouse layer (no merge cost amortized at write time).
  • Classic insert-triggered MATERIALIZED VIEW objects attached to the destination table do NOT fire — MOVE PARTITION is invisible to MV triggers. Any CREATE MATERIALIZED VIEW … TO <events_feed> AS SELECT … FROM <table> will freeze at the rows that existed at MV creation time and silently never update. Use a REFRESHABLE MATERIALIZED VIEW (ClickHouse 23.12+) instead — it polls on a schedule and atomically swaps the target.

Step 2: Find the Correct Connector Version

Always use the latest numbered version tag. Query the connector registry:

flowctl raw get --table connector_tags \
  --query 'documentation_url=eq.https://go.estuary.dev/materialize-clickhouse' \
  --query 'select=image_tag,documentation_url' \
  --output yaml

Use the returned image_tag — never hardcode a version.

Step 3: Help User Complete Prerequisites

Walk the user through prerequisites from the docs loaded in Step 0:

  1. ClickHouse database and user — Create a user with a password

  2. Database permissions — Grant on the target database:

    GRANT ALL ON <database>.* TO <user>;
    
  3. System table permissions — These are NOT covered by the database grant above. The connector needs them for metadata discovery and partition management:

    GRANT SELECT ON system.columns TO <user>;
    GRANT SELECT ON system.parts TO <user>;
    GRANT SELECT ON system.tables TO <user>;
    
  4. Optional: row-level policies — To restrict the user's system table access to only the target database:

    CREATE ROW POLICY estuary_tables ON system.tables FOR SELECT USING database = '<database>' TO <user>;
    CREATE ROW POLICY estuary_columns ON system.columns FOR SELECT USING database = '<database>' TO <user>;
    CREATE ROW POLICY estuary_parts ON system.parts FOR SELECT USING database = '<database>' TO <user>;
    
  5. allow_nullable_key enabled — Required at the MergeTree level (not user profile) because the connector ORDERs BY a Nullable(Int64) column. ClickHouse Cloud has this on already.

    For self-hosted servers, drop a snippet into /etc/clickhouse-server/config.d/:

    <!-- /etc/clickhouse-server/config.d/allow_nullable_key.xml -->
    <clickhouse>
      <merge_tree>
        <allow_nullable_key>1</allow_nullable_key>
      </merge_tree>
    </clickhouse>
    

    Then restart the server. Putting this under <profiles> in users.d/ will not work and causes startup to fail with Setting allow_nullable_key is neither a builtin setting nor started with the prefix 'SQL_'.

  6. Network access — Ensure the ClickHouse server is reachable on the Native protocol port (9440 or 9000) from Estuary

Refer to the docs page for exact SQL commands.

Step 4: Create the Spec File

Build flow.yaml using the config reference from the docs. Minimal required config:

materializations:
  <TENANT>/<PATH>/materialize-clickhouse:
    endpoint:
      connector:
        image: ghcr.io/estuary/materialize-clickhouse:<VERSION>
        config:
          address: "<HOST>:9440"
          credentials:
            auth_type: "user_password"
            username: "<USERNAME>"
            password: "<PASSWORD>"
          database: "<DATABASE>"
          # hardDelete: false           # default — soft-delete via _meta/op
          # advanced:
          #   sslmode: "verify-full"    # default; or "require" / "disable"
          #   no_flow_document: false   # exclude root document column from standard updates
    bindings:
      - source: <TENANT>/<collection-path>
        resource:
          table: "<TABLE_NAME>"
          # delta_updates: false        # default; true to use MergeTree append-only

Important:

  • address must include the port — use 9440 for TLS (default) or 9000 for plaintext.
  • auth_type must be user_password.
  • Setting sslmode: "disable" requires using port 9000.

Step 5: Publish

flowctl catalog publish --source flow.yaml --auto-approve

Step 6: Verify

# Check status
flowctl catalog status <TENANT>/<PATH>/materialize-clickhouse

# View logs
flowctl logs --task <TENANT>/<PATH>/materialize-clickhouse --since 5m | jq -c '{ts, message}'

Status progression:

  1. PENDING — Normal for ~30 seconds during shard assignment
  2. BACKFILLING — Initial data sync from collections
  3. OK — Running normally with real-time updates

Verify rows in ClickHouse using FINAL:

SELECT * FROM <database>.<table> FINAL LIMIT 10;

Without FINAL, you may see duplicate rows or tombstoned rows that haven't been merged yet.

Troubleshooting

"connection refused" or timeout

Cause: ClickHouse not reachable on the Native protocol port

Fix:

  1. Verify port 9440 (TLS) or 9000 (no TLS) is open — not 8123 (HTTP interface, unsupported)
  2. For ClickHouse Cloud, verify the IP allowlist includes Estuary's IPs (see docs)
  3. For self-hosted, check the server's listen_host and tcp_port_secure / tcp_port settings

"authentication failed" or "Wrong password"

Cause: Incorrect credentials or auth type mismatch

Fix:

  1. Verify the username and password
  2. Ensure auth_type is user_password (the only supported value)
  3. Check the user exists: SELECT name FROM system.users WHERE name = '<user>';

"Not enough privileges" on system.columns / system.parts / system.tables

Cause: System table grants were skipped — GRANT ALL ON <db>.* does not cover system tables

Fix: Explicitly grant SELECT on the system tables:

GRANT SELECT ON system.columns TO <user>;
GRANT SELECT ON system.parts TO <user>;
GRANT SELECT ON system.tables TO <user>;

TLS / SSL handshake errors

Cause: SSL mode mismatch with the server's TLS configuration

Fix:

  1. ClickHouse Cloud requires TLS — use sslmode: "verify-full" (default) with port 9440
  2. For self-hosted without TLS, use sslmode: "disable" with port 9000
  3. For self-signed certs, use sslmode: "require" to skip CA verification
  4. Check that the address port matches the SSL mode

Queries return duplicate rows or unexpected deleted rows

Cause: Query is missing the FINAL directive

Fix: ClickHouse merges duplicates in the background — queries must use FINAL to deduplicate at query time:

SELECT * FROM my_table FINAL;

Deletions don't remove rows from the destination

Cause: Soft-delete mode is on by default — deleted rows are kept with _meta/op = 'd'

Fix: Set hardDelete: true in the endpoint config. This makes deletes insert a tombstone row (_is_deleted = 1) that ClickHouse uses to exclude rows from FINAL queries and eventually remove from persistent storage.

"Sorting key contains nullable columns" / Code 44 on CREATE TABLE

Cause: ClickHouse refuses Nullable(...) columns as sort keys by default. The connector ORDERs BY _meta/row_id (Nullable(Int64)).

Fix: Enable allow_nullable_key at the MergeTree level (not user profile). On self-hosted servers, add /etc/clickhouse-server/config.d/allow_nullable_key.xml:

<clickhouse>
  <merge_tree>
    <allow_nullable_key>1</allow_nullable_key>
  </merge_tree>
</clickhouse>

Restart the server. ClickHouse Cloud has this set already. Do NOT put this under <profiles> in users.d/ — it's a MergeTree-level setting and the wrong placement makes the server fail to start with Setting allow_nullable_key is neither a builtin setting nor started with the prefix 'SQL_'.

A MATERIALIZED VIEW over a connector-managed table freezes / doesn't update

Cause: The connector loads data via ALTER TABLE … MOVE PARTITION ID 'all' TO TABLE <table>. MOVE PARTITION is a metadata-only operation and does not fire materialized views subscribed to the destination. Your MV will capture rows that existed at MV-creation time (or any direct INSERTs you make), and silently stop seeing new connector batches.

Fix: Use a REFRESHABLE MATERIALIZED VIEW (ClickHouse 23.12+) instead, which polls on a cron and atomically swaps the target:

CREATE MATERIALIZED VIEW my_agg
REFRESH EVERY 5 SECOND
ENGINE = MergeTree ORDER BY (key)
AS SELECT ... FROM <connector_table> FINAL GROUP BY key;

Inspect status via system.view_refreshes. For sub-second freshness, use a plain VIEW with FINAL and let the query do the work each time.

"Table already exists" or schema conflicts

Cause: Table exists with an incompatible engine or schema

Fix:

  1. Drop and recreate the table — let the connector create it with ReplacingMergeTree (or MergeTree for delta updates)
  2. Use a different table name
  3. Ensure the collection schema is compatible with the existing table

Materialization stuck in PENDING

Wait 30-60 seconds — this is normal during shard assignment. If still stuck:

flowctl logs --task <TENANT>/<PATH>/materialize-clickhouse --since 5m | jq 'select(.level == "error")'

Delta updates: rows accumulating without deduplication

Cause: This is the expected behavior of delta updates — the binding uses MergeTree and appends every store as-is

Fix: This is by design. If you need deduplication, disable delta_updates on the binding and let the connector recreate the table with ReplacingMergeTree.

Related Skills

  • estuary-connector-restart — Pause/restart existing materializations
  • estuary-logs — Deep log analysis
  • estuary-catalog-status — Status checking
slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 197kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 194kUpdated Claude CodeCodexVerifiedInstall