Files

T

Antigravity Agent 93ec7e9056 [G1-Sync] Manual knowledge update

2026-05-09 21:08:02 +09:00

4.6 KiB

Raw Blame History

id, title, category, status, source_trust_level, verification_status, created_at, updated_at, tags, tech_stack, applied_in, aliases

title

Time-series Patterns

메트릭 / 이벤트 / 로그 / IoT = time-series. TimescaleDB (Postgres extension) 가 modern 표준 — 일반 SQL + 시간축 최적화. InfluxDB / Prometheus 도 옵션.

📖 핵심 개념

Hypertable: 시간 기준 자동 파티션.
Continuous aggregate: 실시간 materialized view (1m / 1h / 1d 다운샘플).
Compression: 오래된 chunk 자동 압축 (10-30x).
Retention policy: N일 후 자동 drop.

💻 코드 패턴

TimescaleDB hypertable

CREATE EXTENSION IF NOT EXISTS timescaledb;

CREATE TABLE metrics (
  time TIMESTAMPTZ NOT NULL,
  device_id TEXT NOT NULL,
  cpu DOUBLE PRECISION,
  mem DOUBLE PRECISION
);

SELECT create_hypertable('metrics', 'time', chunk_time_interval => INTERVAL '1 day');

CREATE INDEX metrics_device_time ON metrics(device_id, time DESC);

Insert (대량)

COPY metrics FROM STDIN;
-- 또는 multi-row INSERT
INSERT INTO metrics(time, device_id, cpu, mem)
VALUES ('2026-05-09 10:00', 'd1', 0.5, 0.3),
       ('2026-05-09 10:01', 'd1', 0.6, 0.3);

Query (시간 범위)

-- 최근 1시간
SELECT time_bucket('1 minute', time) AS bucket,
       device_id,
       avg(cpu) AS avg_cpu
FROM metrics
WHERE time > NOW() - INTERVAL '1 hour'
GROUP BY bucket, device_id
ORDER BY bucket;

Continuous aggregate

CREATE MATERIALIZED VIEW metrics_hourly
WITH (timescaledb.continuous) AS
SELECT time_bucket('1 hour', time) AS hour,
       device_id,
       avg(cpu) AS avg_cpu,
       max(cpu) AS max_cpu,
       count(*) AS samples
FROM metrics
GROUP BY hour, device_id;

-- refresh 정책 (실시간으로 따라옴)
SELECT add_continuous_aggregate_policy('metrics_hourly',
  start_offset => INTERVAL '3 hours',
  end_offset => INTERVAL '5 minutes',
  schedule_interval => INTERVAL '5 minutes');

Compression

ALTER TABLE metrics SET (
  timescaledb.compress,
  timescaledb.compress_segmentby = 'device_id',
  timescaledb.compress_orderby = 'time DESC'
);

SELECT add_compression_policy('metrics', INTERVAL '7 days');
-- 7일 지난 chunk 자동 압축

Retention

SELECT add_retention_policy('metrics', INTERVAL '90 days');
-- 90일 지난 chunk 자동 drop

Time-bucket gap fill

SELECT time_bucket_gapfill('1 minute', time) AS bucket,
       device_id,
       locf(avg(cpu)) AS cpu  -- last observation carried forward
FROM metrics
WHERE time > NOW() - INTERVAL '1 hour'
  AND time <= NOW()
GROUP BY bucket, device_id;

Downsampling 단계

raw (1s) → 1분 (cont. agg) → 1시간 → 1일

각 단계는 다른 retention.

InfluxDB (alternative)

# Line protocol
metrics,device=d1 cpu=0.5,mem=0.3 1715238000000000000

// Flux 쿼리
from(bucket: "default")
  |> range(start: -1h)
  |> filter(fn: (r) => r._measurement == "metrics")
  |> aggregateWindow(every: 1m, fn: mean)

Prometheus (메트릭 + 알람)

# scrape config
scrape_configs:
- job_name: api
  static_configs:
  - targets: ['api:9090']

PromQL:

rate(http_requests_total[5m])
histogram_quantile(0.99, sum by (le) (rate(http_request_duration_seconds_bucket[5m])))

🤔 의사결정 기준

데이터	추천
메트릭 + alert	Prometheus + Grafana
IoT / 센서 (장기 보관)	TimescaleDB
트레이딩 / 분석	TimescaleDB / ClickHouse
단순 events	Postgres + 파티션
매우 큰 (PB)	ClickHouse / Druid
짧은 (실시간 30일)	Prometheus / InfluxDB

❌ 안티패턴

단일 테이블 1B+ rows 일반 PG: 인덱스 거대, query 느림.
Index time + device 따로: composite (device, time DESC) 가 best.
Retention 없음: 영원 자라남.
Compression 미사용: 디스크 5-10x 더.
Continuous agg 없음 — raw 매번: 분 단위 query 가 시간.
Time as TEXT: 정렬 안 됨, range 쿼리 느림. TIMESTAMPTZ.
Monitoring DB 에 트랜잭션: HFT app + monitoring 같은 PG = 맥 끊김.

🤖 LLM 활용 힌트

TimescaleDB = Postgres 알면 그대로.
Hypertable + cont agg + compression + retention 4종.
Time-bucket + gapfill + locf 활용.

4.6 KiB Raw Blame History