./present --deck=04 --mode=cinematic

streaming
at the speed of decisions

# Terminal aesthetic — scan lines, CRT vignette, blinking cursor.
audience = "engineers"
tone = "technical, dry-witted, confident"

module 02

when stateful stops scaling.

A walk through three failure modes nobody documented because nobody survived the postmortem.

finding · 03

Dashboards show CPU. The problem is RocksDB compaction. The problem is always RocksDB compaction.

01$ kubectl logs ingest-7d8f --tail=8

02[ok] consumer lag: 82 msgs

03[ok] p50 latency: 18ms

04[warn] compaction backlog: 3.2GB

05[err] rocksdb write stall "L0 too many files"

06--

07$ echo "this is the actual bug"

rocksdb kafka-3.6 k8s page-cache

results · prod · 14d

latency.p99

42ms

// -78% vs. v0

cost.per_1m

$0.18

// -64% vs. v0

recovery.sla

7.2s

// -91% vs. v0

if you can't tail it, you can't debug it. if you can't debug it, you can't ship it.

end-of-deck