ClickHouse

What it is

ClickHouse is a columnar OLAP database. It is not a finance product and ships no market data. It is in this catalogue because "where do I put ten years of ticks" has about four real answers and this is one of them.

The engine shape is what matters. A MergeTree table is stored sorted by its ordering key, and that key is also the primary index — sparse, one mark per granule — so ORDER BY (symbol, ts) turns a query bounded by a symbol and a time window into a binary search over marks and a few granules read. Compression is per column and per codec, and the specialised codecs suit tick data: DoubleDelta for monotonic timestamps and sequence numbers, Delta or Gorilla for slowly drifting prices, T64 for narrow integers, with ZSTD or LZ4 over the top. ClickHouse's own July 2026 write-up of Binance futures data reports 16.2 billion trade and quote rows in 76 GiB on disk, about 19x smaller than the source CSV.

Three more features do the rest. ASOF JOIN matches each trade to the most recent quote at or before its timestamp — the join every slippage and TCA query needs, and one most SQL databases make you hand-roll. AggregatingMergeTree behind a materialized view rolls raw ticks into OHLC or VWAP bars as they arrive, so the bar table is never rebuilt. And TTL rules relocate or expire parts on a clause: TTL ts + INTERVAL 1 YEAR TO VOLUME 'cold' pushes old partitions onto S3 while recent months stay on local disk.

The project is alive on any reading: the repo is not archived, the last commit on master landed on 13 September 2026, and the newest release line is 26.8 LTS, opened on 30 August 2026 and patched on 1 September. Point releases across supported branches ship most weeks.

Pricing

All of the above is in the free build; what the vendor sells is operating it. ClickHouse Cloud is metered, not seated — compute per compute-unit-hour (a unit is 8 GiB RAM and 2 vCPU), metered by the minute and scaled to zero when idle, storage per TB-month of compressed data, which for tick tables is the number that matters. No monthly minimum. New accounts get $300 of credits for 30 days, ending on whichever runs out first, then convert to pay-as-you-go.

The tier figures above are the vendor's own AWS us-east-1 worked examples, not seat prices: change the replica size, the active hours or the volume and the bill moves, and the per-unit rate itself varies by region and cloud.

Data & coverage

None — that is the category. Loading is the work: the s3() and url() table functions read Parquet and CSV in place, ClickPipes pulls from Kafka and object storage on Cloud, and otherwise it is batched inserts from a client.

Integrations

HTTP and native TCP interfaces plus MySQL and PostgreSQL wire-protocol ports, so most SQL tools connect without a driver. Official clients for Python (ClickHouse Connect), Go, Java, JavaScript and Rust, plus JDBC and ODBC. Parquet, Arrow, ORC, CSV and JSONEachRow read and write natively, which is what makes a pandas or Polars round trip cheap. There is a first-party MCP server, and Cloud hosts a remote one beside a console assistant that writes SQL.

Limitations

• No market data, no symbology, no corporate actions. Budget for a feed and for adjustment logic separately.
• Updates and deletes are not what a transactional user means. ALTER TABLE ... UPDATE is an asynchronous mutation that rewrites whole parts; the vendor's own best-practice page is titled "Avoid mutations". Lightweight UPDATE writes patch parts instead, but adds overhead to every later SELECT.
• Deduplication is eventual. ReplacingMergeTree collapses duplicates only when parts merge, "at an unknown time" per the docs, so re-inserting a corrected tick and trusting the engine returns double counts until you add FINAL, which is not free.
• Late ticks insert but do not retract. A bar already materialized through an aggregating view will not recompute when a stale print lands.
• A cluster is real operational work — Keeper, shard and replica layout, part-count and merge monitoring, a backup story. Cloud exists because that is expensive.
• Linux and macOS only for builds and the install script; on Windows it is a Docker story.

Alternatives

QuestDB fits better if you want market-data semantics built in rather than assembled from general primitives; ArcticDB suits a Python-only team that wants versioned DataFrames instead of a SQL server; kdb+ is the desk incumbent and prices accordingly. And if the data fits one disk and one person queries it, Parquet read by DuckDB or pandas costs nothing to operate.

See the rest of the tick data storage category.