perf-agent-lab - automated performance run

JPetStore Needs attention

commit c9a37f22 - Merge pull request #1197 from mybatis/renovate/actions-setup-java-6.x - run 20261018-080254 (UTC) - total 215.8 s - LLM: openai/gpt-oss-20b - findings by the LLM agent

Summary. 1 item(s) need attention; the most important is: Duplicate Order ID causing 500 errors on POST /order/new. 2 small latency difference(s) were within run-to-run noise and are not treated as findings.

Requests

51,577

250 users, 60 s

Throughput

861.86 req/s

p95 latency

38 ms

p50 4 / p99 74

Error rate

0.01%

7 failed

CPU (host)

44.4%

max 84.5%, Tomcat JVM avg 24.0%

Tomcat JVM memory

407 MB

peak RSS

Smoke test

passed

39 req

LLM tokens

20058

Pipeline

CODEPull latest code1.12sok - c9a37f22 Merge pull request #1197 from mybatis/renovate/actions-setup
CODEBuild and start the application41.75sok - build 30.7s, ready after 11.0s
CODESanity (smoke) test20.27sok - last known-good script: 39 req, 0 failed, p95 118.8 ms
AGENTSelf-heal the Locust script0sskipped - smoke test passed - no repair needed
CODERun the load test (250 users, 60 s)60.4sok - 51577 requests, 59 resource samples
CODEAnalyze results0.11sok - p95 38.5 ms, 0 spike(s), 1 error cluster(s), verdict healthy
CODEInspect application and GC logs0.02sok - 945 lines, 2 exception type(s), 117 GC events
AGENTCorrelate with logs, code and recent commits86.07sok - LLM agent with code/log tools (2 noise finding(s) dropped)
CODEBuild report and publish0sok - reports/runs/20261018-080254/report.html

Results

Response time (ms)

Throughput & errors

CPU (%)

Network (KB/s) & Tomcat JVM memory (MB RSS)

Endpoint	Req	Fail	avg	p50	p95	p99	max	vs baseline
POST /account/signon	1041	0	16	8	67	133	159	+229%
GET /account/signoff	1039	0	15	8	59	96	152	+346%
GET /catalog/viewProduct	9262	0	10	4	40	80	118	+327%
GET /catalog/viewItem	9262	0	10	4	39	71	119	+332%
GET /catalog/viewCategory	9263	0	9	4	39	71	110	+333%
GET /cart	2074	0	9	4	36	60	82	+333%
GET /catalog	9263	0	9	4	36	75	139	+306%
GET /cart/addItem	3114	0	9	5	36	62	90	+310%
GET /catalog/searchProducts	2061	0	9	4	36	79	128	+262%
POST /order/new (confirm)	1039	7	10	5	34	60	157	+271%
GET /cart/checkout	1040	0	8	4	32	54	81	+290%
GET /account (sign-on form)	1041	0	8	4	30	74	233	+249%
GET /order/new	1039	0	8	4	30	48	76	+315%
POST /order/new (billing)	1039	0	8	4	30	64	80	+295%

Errors

Endpoint	Status	Message	Count	First / last
POST /order/new (confirm)	500	HTTP 500	7	1s / 23s

Logs

Application log

945 lines scanned - levels: SEVERE 9

HsqlException integrity constraint violation: unique constraint or index violation ; PK_ORDERS table: ORDERS x7 OrderService.insertOrder (OrderService.java:85)
ConnectException Connection refused x1

Servlet.service() for servlet [dispatcher] in context with path [/jpetstore] threw exception [Request processing failed; nested exception is (x7)
Could not contact [localhost:8205] (base port [8205] and offset [0]). Tomcat may not be running.
Error stopping Catalina

JVM garbage collection

117 events - total pause 349.6 ms - max 15.3 ms - full GCs 0 - heap after GC (max) 43 MB

Findings

Duplicate Order ID causing 500 errors on POST /order/new high

7 500 errors in 0.6–23s
HsqlException: PK_ORDERS table: ORDERS

Suspected cause: Concurrent order creation attempts with same ORDERID

OrderService.java:85 insertOrder method

Recommendation: Implement idempotent order creation or pre-check for existing ORDERID before insert (recurring issue)

Improvement plan

#	Improvement	Expected impact	Effort
1	Add pre-insert check or use auto-increment for ORDERID to avoid duplicate key violations	High – eliminates 500 errors	medium

Token usage

LLM tokens this run: 20058. Deterministic steps (git, build, smoke, load, metrics, statistics, log parsing, report) never call the model.

LLM step	Calls	Prompt	Completion
correlate	6	15950	4108

heal skipped - smoke test passed, no repair needed (last time it used ~21860 tokens)

Locust and the application ran on the same machine (a shared CI runner), so absolute numbers are indicative and CPU contention is expected; use this report to compare runs and spot regressions. Load: 250 users for 60 s after a 39-request smoke test.