Maven DF: match by known repository base instead of enumeration, when available
Summary
Follow-up to !4291 (closed), proposed by @arpitgogia in this review comment. Credit to him for the idea and the writeup this issue is based on.
!4291 fixes mirror/registry path-prefix matching by enumerating every possible group reading of the path (dropping one leading segment at a time) and checking each against policy. That's correct and necessary as a general fallback, but when we already know the repository's base URL — because the caller configured it in their own settings.xml — we don't need to guess at all: stripping the known base off the path gives the exact group, with one policy call instead of up to ten, and no risk of a shorter enumerated reading colliding with an unrelated real package.
The idea
<mirror>
<id>corp</id>
<mirrorOf>*</mirrorOf>
<url>https://nexus.corp/repository/maven-public/</url>
</mirror>https://nexus.corp/repository/maven-public/org/apache/logging/log4j/log4j-core/2.14.1/log4j-core-2.14.1.jar
└────────────── known base, strip it ─────┘└──────── group ────────┘Stripping the known base gives org.apache.logging.log4j:log4j-core directly. Treating Maven Central's own base as "just another known base" gives the same one-call result for unmirrored builds too. Enumeration would then only run for bases we don't recognize.
Why this is worth doing, not just cheaper
- One call per artifact instead of one per group segment. Directly reduces the cost
!4291's own description flags as an accepted tradeoff (up to a 2.84x increase in policy calls). - Removes a real false-positive class, not just a theoretical one. Enumeration can generate a short reading that collides with an unrelated real package — e.g. a rule on the old relocated
dom4j:dom4jalso blocks the currentorg.dom4j:dom4jeven from plain Central today, and a rule onfoo:barwould also blockcom.foo:barbehind a mirror. Knowing the real base eliminates this category of mismatch entirely rather than reducing how often it occurs. - No dependence on group depth, so
mavenMaxGroupDepth(the cap added to bound enumeration's worst case) becomes irrelevant for any request behind a known base.
What exists already
- The proxy already has the request's host:
serveTunnelsetsreq.URL.Host = authority(proxy.go).MavenMatchercurrently reads onlyreq.URL.Path. !4289already locates the settings file Maven will read (-sinMAVEN_ARGS, the command line,.mvn/maven.config,~/.m2/settings.xml). Reading<mirrors>and profile<repositories>/<pluginRepositories>out of that file is a small step on top of locating it.- One wiring detail to solve: in
!4289,proxy.New(manager.Matcher(), …)runs beforeSetArgsinpm.go'sRun(). The matcher needs the resolved bases earlier than it currently gets the args, or needs to resolve them lazily.
What it can't replace
Enumeration has to remain as a permanent fallback — this is additive, not a replacement:
- Repositories declared in the project POM, a parent POM, or a dependency's POM aren't in
settings.xmlat all. mirrorOfpatterns likeexternal:*or!some-repolet specific repositories bypass a mirror; this issue doesn't take on evaluating those patterns, only stripping URLs we already have.- Global settings (
$MAVEN_HOME/conf/settings.xml,-gs) aren't read by!4289yet — see also the-gsgap already found empirically in that investigation, where the firewall didn't engage at all under-gsagainst released glab. ${env.…}URLs and profile activation add their own resolution logic.- Gradle shares this matcher but has no settings file at all — its repositories live in build scripts.
The unknown-base path must never fall back to the old, unprefixed-only parse — only to !4291's enumeration.
Proposed scope
Known base → exact strip (one call). Unknown base → !4291's enumeration (unchanged). A rough PoC of the stripping mechanism itself (no <mirrorOf> evaluation, no POM awareness) exists on 637331-combined-fix-poc for reference.
Depends on
- !4291 (closed) (Gap 1 — path-prefix enumeration, the fallback this builds on)
- !4289 (closed) (Gap 2 — settings-file discovery, the source of known bases)
Not expected to block either of those landing first.
Relates to gitlab-org/gitlab#637331.