Skip to content

feat: complete ssrf-guard demo matrix (httpclient5-demo + native-image-demo)#61

Merged
jlc488 merged 2 commits into
mainfrom
feat/ssrf-guard-demo-matrix-completion
May 23, 2026
Merged

feat: complete ssrf-guard demo matrix (httpclient5-demo + native-image-demo)#61
jlc488 merged 2 commits into
mainfrom
feat/ssrf-guard-demo-matrix-completion

Conversation

@jlc488
Copy link
Copy Markdown
Contributor

@jlc488 jlc488 commented May 23, 2026

Summary

Two new demos closing out the ssrf-guard module coverage on devslab-examples — every module published to Maven Central now has a runnable demo or a transitive home in one.

Demo What it shows
ssrf-guard-httpclient5-demo The only ssrf-guard module that had no runnable demo before. Apache HttpClient 5's DnsResolver + RedirectStrategy plug-in points, with SafeDnsResolver closing the TOCTOU window between URL validation and Socket.connect().
ssrf-guard-native-image-demo End-to-end proof that v3.1.0's RuntimeHintsRegistrar entries let a Spring Boot app ./gradlew nativeCompile successfully — no consumer-side reflect-config.json required.

What changes

  • + ssrf-guard-httpclient5-demo/ (16 files) — full Spring Boot scaffold mirroring ssrf-guard-okhttp-demo, smoke tests for the 3 DNS-gate block paths
  • + ssrf-guard-native-image-demo/ (16 files) — Spring Boot + org.graalvm.buildtools.native plugin, 12-pattern attack catalog, JVM-mode smoke tests, README documenting the nativeCompile workflow
  • Top-level README.md + README.ko.md — new rows for both demos, with badges

Verification

Both demos ./gradlew build (incl. tests) passes locally against live Maven Central artifacts:

Demo Time Outcome
ssrf-guard-httpclient5-demo 33s ✅ BUILD SUCCESSFUL
ssrf-guard-native-image-demo 1m 21s ✅ BUILD SUCCESSFUL

Native-image build (./gradlew nativeCompile) is not run in CI — it's documented as a local verification step for native-image users. CI's JVM build exercises the same processAot codegen path indirectly via the test suite.

Coverage matrix (now complete)

Module Demo
ssrf-guard-restclient / -resttemplate / -webclient ssrf-guard-demo (combined 3-in-1)
ssrf-guard-feign ssrf-guard-feign-demo
ssrf-guard-okhttp ssrf-guard-okhttp-demo
ssrf-guard-jdkhttp ssrf-guard-jdkhttp-demo
ssrf-guard-springai ssrf-guard-springai-demo
ssrf-guard-langchain4j ssrf-guard-langchain4j-demo
ssrf-guard-httpclient5 ssrf-guard-httpclient5-demo (new)
ssrf-guard meta + AOT hints ssrf-guard-native-image-demo (new)
ssrf-guard-core / -llm covered transitively in every demo above

Test plan

  • Local ./gradlew build passes for both demos
  • CI verifies both demos (CI build matrix detects new directories automatically per existing convention)
  • Reviewer can ./gradlew bootRun on either demo and try the curl snippets from the README

Note on parallel branch work

feat/api-log-demos is a separate parallel effort (api-log demos — api-log-jpa-demo, -mybatis-demo, -r2dbc-demo show up as untracked when both branches are pulled together). This PR scopes only the ssrf-guard demos and intentionally doesn't touch api-log demos.

…demo

Two new demos closing out the ssrf-guard module coverage:

ssrf-guard-httpclient5-demo
---------------------------
The only ssrf-guard module without a runnable demo before this. Apache
HttpClient 5 plugs the SSRF policy in at DNS-resolution time rather
than URL-parse time — a different shape from the other client demos.

- SafeDnsResolver (DnsResolver SPI): rejects hosts outside the whitelist,
  filters private/loopback/link-local/cloud-metadata IPs out of the
  resolved set. The InetAddress[] it returns is what HttpClient hands to
  Socket.connect(), closing the TOCTOU window.
- SafeRedirectStrategy (RedirectStrategy SPI): same DNS gate runs on
  every redirect hop.
- The module ships its own Spring autoconfig
  (SsrfGuardHttpClient5AutoConfiguration), so the demo's main() needs
  zero wiring code. Outside Spring, five lines on HttpClients.custom().

The README acknowledges the trade-off honestly: this approach doesn't do
URL-parse-time gates (scheme restriction, userinfo rejection) but does
close all the practical attack vectors at the resolver, and the
TOCTOU-closing property is the main selling point.

ssrf-guard-native-image-demo
----------------------------
End-to-end proof that ssrf-guard 3.1.0's RuntimeHintsRegistrar entries
let a Spring Boot app `./gradlew nativeCompile` without writing any
reflect-config.json for our types.

- Pulls kr.devslab:ssrf-guard:3.1.0 (the meta artifact — -restclient
  + -httpclient5 + -core, hints come along).
- Applies org.graalvm.buildtools.native plugin.
- Exposes /fetch + /attacks (a 12-pattern catalog).
- Surfaces "runtime": "jvm" vs "graalvm-native" so visual inspection
  confirms which build you're hitting.

CI builds JVM mode only (native-image is too slow / memory-heavy for
every PR). README documents the local verification flow:
- ./gradlew processAot (fast ~10s) to verify hints register
- ./gradlew nativeCompile (3-8 min) to produce a working binary

Top-level README + README.ko.md updated with rows for both new demos.

Verified: ./gradlew build passes for both demos in JVM mode against the
live Maven Central kr.devslab:ssrf-guard*:3.1.0 artifacts (httpclient5
build 33s, native-image build 1m21s).
Same Windows-clone gotcha that bit ssrf-guard-langchain4j-demo earlier
(devslab-examples/#52 → fixup commit 0971c1b). The gradlew scripts in
ssrf-guard-httpclient5-demo and ssrf-guard-native-image-demo were
committed as 100644 (no execute bit), so Linux CI failed with:

  /home/runner/.../gradlew: Permission denied
  Process completed with exit code 126.

Re-stage with chmod=+x → 100755. No content change.
jlc488 added a commit that referenced this pull request May 23, 2026
CI failed with `./gradlew: Permission denied` (exit 126) on all three
new demos. The agents created the wrappers via `cp -r` on Windows,
where the filesystem doesn't track executable bits, so git stored them
as 0644. Linux CI runners then refused to execute them.

`git update-index --chmod=+x` flips the index mode to 100755 without
changing file content. Same fix as #61 used (commit 2dbf4a4).
@jlc488 jlc488 merged commit 1776ba4 into main May 23, 2026
3 checks passed
jlc488 added a commit that referenced this pull request May 25, 2026
…stcontainers IT (#62)

* feat: add api-log demo set (jpa + mybatis + r2dbc) with PostgreSQL Testcontainers IT

Covers all three persistence backends of the api-log starter
(kr.devslab:api-log-{jpa,mybatis,r2dbc}:3.0.0) so a reader can pick
the demo that matches their stack and `./gradlew bootRun` immediately.

## Demos added

  api-log-jpa-demo/       Spring MVC + JPA + Postgres        (blocking)
  api-log-mybatis-demo/   Spring MVC + MyBatis + Postgres    (blocking)
  api-log-r2dbc-demo/     WebFlux + R2DBC + Postgres         (reactive)

## Common design — self-loopback

Each demo exposes three controllers in the same app:

  /upstream/widgets/{id}         the "service being called" — returns a
                                 fake Widget, with id=999 forcing a 5xx
                                 to exercise the error path
  /client/widgets/{id}           calls upstream via RestApiClientUtil
                                 (or ReactiveApiClientUtil for r2dbc) so
                                 the call gets logged into api_log
  /client/widgets/with-request-id/{id}
                                 same shape but passes an explicit
                                 requestId via the core
                                 send(HttpMethod, ApiRequest) overload —
                                 demonstrates the retry-correlation API
  /api-log/recent                reads the api_log table (top 20 by
                                 timestamp DESC)
  /api-log/by-request/{rid}      reads all rows for one requestId
                                 (lifecycle: INITIATED → SUCCESS / ERROR)
  /api-log/by-event/{type}       reads all rows for one event type

Self-loopback means no external dep — `docker compose up -d db &&
./gradlew bootRun && curl localhost:8080/client/widgets/123` is the full
demo. Reader can immediately see INITIATED + SUCCESS pairs in
/api-log/recent after the async event listener flushes.

## Tests — Testcontainers + @Serviceconnection

Each demo ships an integration test (ApiLogLifecycleIT) that spins up
postgres:16-alpine via Testcontainers, makes real HTTP self-loopback
calls (RestClient.create / TestRestTemplate / WebTestClient depending
on the demo's stack), and asserts on the api_log rows that the async
listener writes. Awaitility polls past the listener's async window.

Five tests per demo:
  1. happy GET path → INITIATED + SUCCESS in api_log
  2. error path (id=999) → INITIATED + ERROR
  3. POST body preserved in api_log.payload (JSONB)
  4. explicit requestId correlation via /with-request-id/{id}
  5. schema initialized on boot (table exists, query succeeds)

15 IT tests total across the three demos. CI builds them automatically
(the workflow's detect step picks up new demos by the presence of
build.gradle.kts — no workflow edits needed).

## Backend-specific notes

JPA:
  - spring-boot-starter-data-jpa + api-log-jpa
  - Reader uses ApiLogRepository (the starter publishes it as a
    Spring Data repository — extends JpaRepository<ApiLogEntity, Long>)
  - spring.jpa.hibernate.ddl-auto=none — api-log starter's
    ApiLogJpaSchemaInitializer creates the api_log table itself
    (api.log.schema.management=BUILTIN)

MyBatis:
  - mybatis-spring-boot-starter:4.0.1 + api-log-mybatis
  - Reader uses the bundled ApiLogMapper for findByRequestId,
    plus a custom ApiLogQueryMapper (xml) for findRecent / findByEvent
    since the starter's mapper doesn't expose those
  - @MapperScan scoped to the demo package only (the starter's mapper
    is registered by its own auto-config — scanning it twice conflicts)

R2DBC:
  - spring-boot-starter-webflux + spring-boot-starter-data-r2dbc +
    api-log-r2dbc
  - Reader uses DatabaseClient (not a Spring Data R2DBC repository —
    the api-log r2dbc backend doesn't ship one, keeping its dep
    footprint minimal). Cast JSONB columns to ::text so they bind
    cleanly to String fields on a public ApiLogView record.
  - Schema initializer is opt-in via api.log.r2dbc.schema.enabled=true
    (separate property from the JDBC backends because reactive init
    runs on ConnectionFactory not DataSource — per ApiLogProperties
    javadoc)
  - Both r2dbc-postgresql AND jdbc postgresql drivers in the build —
    the latter is what Testcontainers' @Serviceconnection prefers for
    rewiring the connection.

## Root README updates

Added an "### api-log" section after the ssrf-guard section in both
README.md and README.ko.md, with the three new demos linked from a
single table. Matches the existing table convention (Maven Central
shields.io badge with verbose label) — a follow-up that strips the
labels to the default "Maven Central" form (per devslab-examples#59)
would touch all rows at once.

## Sibling work in flight

devslab-examples#61 adds two more ssrf-guard demos
(ssrf-guard-httpclient5-demo + ssrf-guard-native-image-demo) on a
parallel branch. The two PRs are independent — they touch different
directories. README.md will conflict on the section ordering after both
land; whichever merges second rebases.

* fix(api-log demos): mark gradlew executable in git index

CI failed with `./gradlew: Permission denied` (exit 126) on all three
new demos. The agents created the wrappers via `cp -r` on Windows,
where the filesystem doesn't track executable bits, so git stored them
as 0644. Linux CI runners then refused to execute them.

`git update-index --chmod=+x` flips the index mode to 100755 without
changing file content. Same fix as #61 used (commit 2dbf4a4).

* fix(api-log demos): downgrade to Spring Boot 3.5.6 baseline

CI failed with BeanTypeDeductionException -> ClassNotFoundException
during Spring context load on all three demo ITs. Root cause: api-log
3.0.0 is built against Spring Boot 3.5.6 per the Spring-major-aligned
versioning policy (devslab-kr/.github/.github/VERSIONING.md — lib
major = SB major, so api-log 3.x = SB3 line). My demos were on SB4.0.6
and the runtime classpath had api-log's compiled bytecode referencing
SB3.5 classes that don't exist in SB4 — gradle resolution upgraded the
spring-boot-* artifacts to 4.0.6 but didn't paper over the API drift.

Changes per demo (build.gradle.kts):

  Plugin                4.0.6 -> 3.5.6
  Test starters         removed SB4-only -webmvc-test / -webflux-test /
                        -resttestclient modules (MockMvc / WebTestClient /
                        TestRestTemplate all ship in plain
                        spring-boot-starter-test on SB3)
  Testcontainers BOM    removed (relying on Spring Boot 3.5.6's managed
                        Testcontainers versions — same approach as
                        easy-paging-postgres-demo + easy-paging-reactive-demo)
  Testcontainers names  testcontainers-postgresql -> postgresql,
                        testcontainers-junit-jupiter -> junit-jupiter
                        (Testcontainers 1.x naming convention, what SB3
                        BOM pins)
  MyBatis starter       mybatis-spring-boot-starter:4.0.1 -> :3.0.4
                        (SB3-compatible line; matches what PageHelper +
                        Spring Boot 3 BOM expects)
  R2DBC extras          added testcontainers:r2dbc (the R2DBC
                        @Serviceconnection bridge for SB3 — matches
                        easy-paging-reactive-demo)
  Unchanged             api-log 3.0.0 deps, awaitility 4.2.2, R2DBC +
                        JDBC postgres drivers, dependency-management
                        1.1.7, JDK 21 toolchain, junit-platform-launcher

Changes per IT (imports):

  PostgreSQLContainer   org.testcontainers.postgresql.PostgreSQLContainer
                        -> org.testcontainers.containers.PostgreSQLContainer
                        (Testcontainers 1.x class location)
  Container generic     raw PostgreSQLContainer -> PostgreSQLContainer<?>
                        field type + new PostgreSQLContainer<>("...")
                        (1.x self-typed generic, diamond form)
  @AutoConfigureWebTestClient (r2dbc only)
                        org.springframework.boot.webtestclient.autoconfigure
                        -> org.springframework.boot.test.autoconfigure.web.reactive
  TestRestTemplate (mybatis only)
                        org.springframework.boot.resttestclient
                        -> org.springframework.boot.test.web.client
  @LocalServerPort     no change (lives at
                        org.springframework.boot.test.web.server in both
                        SB3.5.6 and SB4 — spec mid-flight said otherwise,
                        agents caught it via compile error)

Test logic itself unchanged (5 tests × 3 demos = 15 IT tests). Main-
source files (controllers, Widget record, ApiLogReader/View, MyBatis
mapper xml, application.yml, docker-compose.yml, READMEs) all
source-compatible between SB3 and SB4 — only build pin + a handful of
test imports needed to move.

Compile verified clean on all 3 demos. Awaiting CI for the runtime path.

Follow-up: once api-log ships a 4.x line (the VERSIONING.md policy
schedules it for whenever the project decides to add SB4 support), add
api-log-{jpa,mybatis,r2dbc}-sb4-demo siblings — same pattern as the
easy-paging dual-line setup.

* fix(api-log demos): self-loopback URL uses runtime local.server.port

The 3 CI test failures (happyGet / postBody / explicitRequestId — all
the success-path tests) were a real bug, not a flake: ClientController
read `${api-log-demo.upstream-base-url}` which expanded to
`http://localhost:8080` at @value injection time. The integration test
boots on RANDOM_PORT — e.g. 54321 — so:

  Test → http://localhost:54321/client/widgets/123     OK
  ClientController → http://localhost:8080/upstream/widgets/123  CONN REFUSED → 500

errorPath and schemaInit passed for the wrong reason — errorPath asserts
an ERROR row shows up regardless of cause (connection-refused counts),
and schemaInit never touches HTTP.

Fix: don't hardcode the port in config. ClientController now injects
Environment and builds the upstream URL at request time from
`local.server.port` (which Spring Boot sets for BOTH bootRun and
RANDOM_PORT after the embedded server binds). One code path, one
behaviour, no test-specific override.

Removed `api-log-demo.upstream-base-url` from all 3 application.yml
since nothing reads it anymore; comment now points users at
ClientController for the "swap upstream URL for production" instruction.

All three demos compile clean. The bug existed in the original SB4
version too — only surfaced after the SB3 downgrade got past the
ContextLoader exception and tests could actually run.

* chore(api-log demos): bump 3.0.0 -> 3.0.1 (HTTP client Content-Type + PATCH fixes)

api-log v3.0.1 published to Maven Central
(devslab-kr/api-log#4 + #5 merged, tag pushed, release workflow succeeded
at 16:39Z, all four artifacts indexed). 3.0.1 fixes two bugs in
RestApiClientUtil / ReactiveApiClientUtil that the previous PR #62 CI
runs surfaced:

  1. Content-Type missing on POST/PUT/PATCH body -> upstream returned
     415/500 (the postBodyIsPreservedInPayloadColumn failures across all
     three demos).
  2. PATCH method unsupported because SimpleClientHttpRequestFactory
     wraps java.net.HttpURLConnection (which rejects PATCH).

Bump scope: one line per demo build.gradle.kts (well, two — core +
backend), three demos = 6 lines. No source changes needed — the bugs
were entirely in the starter, the demo code was already calling the
right APIs.

Expecting all 15 IT tests (5 per demo) to pass on this run now that the
starter is fixed.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant