log-forensicslisted
Install: claude install-skill polytheseus/sre-skills
# Log forensics
Logs are too big to read. Use them like a database: **count, bucket,
compare, then read a few lines**. Read only the lines your numbers point to.
## Ground rules
- Never `cat` an unknown-size log into the conversation. Check the size first
(`wc -l`, `ls -lh`) and sample with `head`, `tail`, `shuf -n`.
- Logs often contain PII, tokens and secrets. Quote the minimum needed, and
redact tokens, emails and IPs in anything you summarise.
- Work in a scratch directory. Don't modify the original files.
## 1. Understand the format
```bash
ls -lh <files>; wc -l <file>
head -3 <file>; tail -3 <file>
```
Identify the timestamp field and format, the level or severity field, and
any request/trace ID. For JSON lines, list the keys:
```bash
head -200 <file> | jq -r 'keys[]' 2>/dev/null | sort | uniq -c | sort -rn
```
For compressed or rotated logs use `zcat`/`zgrep`. For journald, use
`journalctl -u <unit> --since "<time>" --until "<time>" -o short-iso`.
For containers, use `kubectl logs --since=…` or `docker logs --since …`.
## 2. Bucket over time
Count events per minute (adjust the cut to your timestamp format) and look
for the step change:
```bash
# Plain text, ISO timestamps at line start: 2026-01-01T12:34:56...
grep -E 'ERROR|error|Exception' <file> | cut -c1-16 | sort | uniq -c
# JSON lines
jq -r 'select(.level=="error") | .time[0:16]' <file> | sort | uniq -c
```
Find the **first bad minute**. Everything after it is effect. The cause is
usually in the m