장애를 설명하지 않는다.직접 만들고, 숫자로 목격한다.

STAMP 05 · lost-update repair workshop

같은 1000을 읽은 두 debit이 왜 800이 되지 않는지 trace에서 잡는다

Fixed incident: balance 1000 · worker A/B each debit 100 · serial business invariant 800 · PostgreSQL READ COMMITTED.

SOURCE

exact backend strategy class excerpt

SERVER

strategy ID → precompiled Spring path

NOT CLAIMED

browser Java compilation or arbitrary execution

whole system · evidence stays attached to its layer

한 HTTP request 안에서 두 DB transaction이 갈라졌다가 다시 합쳐진다

01
Browser
source + prediction
02
Next gateway
allowlist + rate/gate
03
Tomcat
request coordinator
04
JVM workers
2 transaction threads
05
Hikari
bounded JDBC sessions
06
PostgreSQL
row · version · lock

JVM CPUNOT OBSERVED

scheduler time · saturation

JVM MEMORYNOT OBSERVED

heap · allocation · GC

NETWORKNOT OBSERVED

RTT · request/response bytes

PG RESOURCESNOT OBSERVED

CPU · buffers · disk reads

PostgreSQL lock wait is separately MEASURED through pg_stat_activity + pg_blocking_pids. The four machine-resource groups above remain NOT OBSERVED.

single exact strategy-excerpt editor · src-d4fced0b

Read then absolute write

MATCHED BACKEND CLASS

Review question: What happens when both transactions calculate the same absolute balance before either commit is visible?

The displayed class excerpt is normalized directly against LostUpdateStrategies.java and stays in the browser. Only its derived strategy ID is sent; the server executes that precompiled class in the Spring image.