mkpipe-loader-clickhouse 0.10.0

Clickhouse loader for mkpipe.

mkpipe-loader-clickhouse

ClickHouse loader plugin for MkPipe. Writes Spark DataFrames into ClickHouse tables using the native clickhouse-spark connector, which uses ClickHouse's binary HTTP protocol for efficient columnar inserts.

Documentation

For more detailed documentation, please visit the GitHub repository.

License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.

---

Connection Configuration

connections:
  clickhouse_target:
    variant: clickhouse
    host: localhost
    port: 8123
    database: target_db
    user: default
    password: mypassword

---

Table Configuration

pipelines:
  - name: pg_to_clickhouse
    source: pg_source
    destination: clickhouse_target
    tables:
      - name: public.events
        target_name: stg_events
        replication_method: full
        batchsize: 50000

---

Write Strategy

Control how data is written to ClickHouse:

      - name: public.events
        target_name: stg_events
        write_strategy: upsert       # append | replace | upsert
        write_key: [id]              # required for upsert

Strategy	ClickHouse Behavior
append	Insert via ClickHouse Spark connector (default for incremental)
replace	Drop and recreate table, then insert (default for full)
upsert	Creates table with ReplacingMergeTree engine using write_key as ORDER BY. ClickHouse deduplicates rows with the same key on background merges.

Note: ClickHouse does not support SQL MERGE statements. Upsert semantics are achieved via ReplacingMergeTree, which deduplicates asynchronously during compaction. Use FINAL in queries to get deduplicated results.

---

Write Parallelism & Throughput

ClickHouse loader inherits from JdbcLoader. Two parameters control write performance:

      - name: public.events
        target_name: stg_events
        replication_method: full
        batchsize: 50000        # rows per JDBC batch insert (default: 10000)
        write_partitions: 4     # coalesce DataFrame to N partitions before writing

How they work

• batchsize: number of rows buffered before sending a single INSERT to ClickHouse. ClickHouse benefits greatly from large batches — use 50,000–500,000 for best throughput.
• write_partitions: calls coalesce(N) on the DataFrame before writing, reducing the number of concurrent JDBC connections. Useful when you have many Spark partitions and want to limit load on ClickHouse.

Performance Notes

• ClickHouse is optimized for large bulk inserts. batchsize is the most impactful parameter — increase it as much as your driver memory allows.
• Avoid many small write_partitions (e.g. 1) as it reduces parallelism. A value of 4–8 balances load and throughput.
• ClickHouse's MergeTree engine merges parts in the background; very frequent small inserts create many parts and degrade query performance. Prefer fewer large batches.

---

All Table Parameters

Parameter	Type	Default	Description
name	string	required	Source table name
target_name	string	required	ClickHouse destination table name
replication_method	full / incremental	full	Replication strategy
batchsize	int	10000	Rows per JDBC batch insert
write_partitions	int	—	Coalesce DataFrame to N partitions before writing
write_strategy	string	—	append, replace, upsert
write_key	list	—	Key columns for upsert (used as ReplacingMergeTree ORDER BY)
dedup_columns	list	—	Columns used for mkpipe_id hash deduplication
tags	list	[]	Tags for selective pipeline execution
pass_on_error	bool	false	Skip table on error instead of failing