Time Series Database Overview

The storage tier of the data lake is Aster TSDB, optimized for industrial time series workloads (apache-cassandra-timeseries). It is an in-house fork of Apache Cassandra cassandra-6.0; the CQL syntax and disk format are identical to the upstream — it reads existing 6.0 data as-is, and new features work only when explicitly enabled.

This is not the "Time Series Engine (TSE)"

These names are similar and are the most common source of confusion. What they store is different.

	Time Series Database (this document)	Time Series Engine (TSE)
What it stores	Sensor values — tag points from the field (pp.tm_tag_point)	Metrics — the system's own indicators, such as PLANTPULSE.JMX.* and PLANTPULSE.EDGE.*
Nature	General-purpose time series database	Metrics-only
Port	9042 (CQL)	7800 (plaintext) · 7801 (TLS)
Query	CQL — with built-in functions like time_bucket_gapfill	TQL
Documentation	This section	Time Series Engine

"What was the equipment temperature?" goes here. "How many records did the collector process yesterday?" goes to the engine.

The engine also writes to the same Aster TSDB ring, but its keyspace and schema are different, so metrics stored in the engine will not appear even if you query the tag tables with CQL.

What Improves

Field time series have their own characteristics. Tags accumulate endlessly at second-level granularity, must be retained for years due to regulations, an edge that was disconnected may push days' worth of data at once (late loading), and queries are almost always “this tag, this time range”. Stock Cassandra can do this, but compression, retention, and aggregation all become the application's responsibility. This fork moves that responsibility into the database itself.

What you want to do	What this database does	Documentation
Handle bucketing, aggregation, and interpolation on the server	21 CQL functions + gap filling	Time Series Functions
Automatically compress old data	Tiered storage — queries remain unchanged	Tiered Storage
Delete old data entirely after the retention period	Time Series Compaction (TSCS)	Time Series Compaction
Search log/event bodies by substring	Full-text search — including Korean	Full-Text Search

Enabling All Three Together

쓰기 ──▶ [ 핫 구간 ]           최근 데이터. 평소대로 행으로 저장·질의
              │
              │  hot_window 를 지나면 (계층형 저장)
              ▼
         [ 콜드 청크 ]          컬럼 지향으로 압축. SELECT 가 알아서 합쳐 읽는다
              │
              │  cold_window / retention 이 지나면
              ▼
           통째로 삭제          컴팩션 없이 창 단위로 (TSCS)

TSCS and tiered storage are independent of each other. You can enable one or both; the lower-risk option is to enable TSCS first.

Measured Results — Production Table Shape, 20 Million Rows

tm_tag_point (8 general columns), host 234, chunk format v4.

Item	Before Tiering	After Tiering	Effect
Storage size	237.8 MB	33.3 MB	7.1× smaller (11.9 → ~1.7 bytes per row)
count(*) (40,000-row partition)	303 ms	50 ms	6.1× faster
time_bucket + aggregation	150–270 ms	31–64 ms	3–6× faster
Gap filling (locf · interpolate)	162–284 ms	30–53 ms	5.4× faster
Hourly average of 90 tags (3.6 million rows)	14.9 s	4.2 s	3.5× faster

Compression also speeds up queries

Usually compression trades speed for space, but here all measured queries were as fast or faster — because fewer bytes are read. Even scans without a range limit stop when LIMIT fills, as they decompress chunks in query order, so they don't get slower.

Two Caveats

• Full-scan aggregation without a partition key returns incorrect values on tiered tables. That path does not merge chunks.
• The space savings depend on the table shape. A minimal shape with no constant columns and no all-null columns (a single high-entropy double) will not see the figures above. Re-measure on your own table before deciding.

Where to Find It

Go to $PP_HOME/plantpulse-storage/db/cassandra then apache-cassandra-timeseries-<version>.jar. The version of the running process is checked as follows.

docker exec plantpulse-datalake nodetool version

Version, source, and update procedure are in the operator documentation, not this manual.

---

See Also

What you want to do	Documentation
Use bucketing, gap filling, and interpolation	Time Series Functions
Move old values to cheaper storage	Tiered Storage
Understand compaction	Time Series Compaction (TSCS)
Search bodies	Full-Text Search
Query directly with CQL	Data Query

Looking for metrics (system indicators)? This isn't the place → Time Series Engine