Skip to main content

Hierarchical Storage

Old windows are compressed into column-oriented chunks and moved to a side table, and SELECT reads them by automatically merging them. Applications don't need to know whether data is compressed.

쓰기 ──▶ [ 핫 구간 ]  ── hot_window 경과 ──▶ [ 콜드 청크 ]  ── cold_window 경과 ──▶ 삭제
           행 저장                              <테이블>__chunks
                    ◀────── SELECT 이 알아서 합친다 ──────▶

Configuration​

The policy is set as a JSON string in the timeseries_tiering key of the table's extensions map.

ALTER TABLE pp.tm_tag_point WITH extensions = {
  'timeseries_tiering': '{"hot_window":"2d","chunk_window":"1d","cold_window":"3650d","interval":"1h"}'
};
KeyRequiredDefaultDescription
hot_windowRequired—Rows younger than this age are left untouched. Must be chunk_window or higher
chunk_window1hLength of the re-encoding window. 1 chunk = 1 tag × 1 window
cold_windowNoneChunks older than this are deleted entirely. Must be greater than hot_window
consistencyLOCAL_QUORUMCL used by the re-encoder. Quorum series only
interval5mRe-encoding interval for this table

Duration syntax follows <양의 정수><m|h|d>. Unknown keys or rule violations (hot_window < chunk_window, cold_window <= hot_window, CL below the quorum, etc.) are rejected when ALTER TABLE occurs — not silently ignored.

No codec choice

The column type determines the encoding. double supports only ALP/ALP-RD, columns with no value changes are handled as CONSTANT, and all-null columns as ALL_NULL, in O(1) regardless of row count. Including a codec key in the policy JSON points to that key and is rejected.

Release​

ALTER TABLE pp.tm_tag_point WITH extensions = {};

Only new encoding stops; data already in chunks continues to appear merged.

Operations​

# 지금 즉시 한 사이클 (interval 무시)
nodetool retier <keyspace> <table>

# 정책이 걸린 모든 테이블의 상태
nodetool tieringstatus

The same content can be viewed via CQL.

SELECT * FROM system_views.timeseries_tiering;
Less hierarchical when tags_skipped is not 0

This means the last completed cycle didn't encode everything up to the cutoff — so it's not as compressed as other counters suggest. retier fails with a non-zero exit code if any tag was skipped. Original rows are untouched, so no data is lost. Check the logs for the cause and re-run.

Statistics are node-local · in-memory and based on the last completed run — they reset on restart. On multiple nodes, each node re-encodes only the tags in its own primary token range, dividing the work.

Things to Know​

  • Cold data is immutable. Deletes are rejected for intervals already in chunks.
  • Late-arriving rows are merged column by column.
  • 🔴 Full-scan aggregations without a partition key return incorrect values on hierarchical tables. That path does not merge chunks. Send such queries to Spark SQL.
  • The payload format is v4 only. If an old removed format is encountered, SELECT will fail, not silently skip — better than continuing to return results missing only historical data.

See Also​