Dominik Schmidt created TIKA-4864:
-------------------------------------
Summary: Default plugins directory resolution depends on the
working directory
Key: TIKA-4864
URL: https://issues.apache.org/jira/browse/TIKA-4864
Project: Tika
Issue Type: Improvement
Reporter: Dominik Schmidt
The 4.0.0 docker image (and any classpath-launched tika-server) loses all pipes
plugins when the process working directory is not the install directory. The
server starts fine, but the first parse fails:
org.apache.tika.exception.TikaConfigException: Unknown fetcher type:
'file-system-fetcher'
for instance id '__tika-server'. Available types: []
Reproduce with the released image; the only difference is the working directory:
docker run -p 9998:9998 apache/tika:4.0.0-full # parses fine
docker run -w /somewhere -p 9998:9998 apache/tika:4.0.0-full # first parse
fails as above
This is not a synthetic case: any orchestrator that sets the container working
directory hits it, for example Woodpecker CI service containers (workspace as
workdir), Kubernetes workingDir, or compose working_dir.
Cause, in TikaServerProcess.resolveDefaultPluginsDir() (introduced with
TIKA-4682):
1. It looks for plugins next to "the running jar", taken from the CodeSource of
TikaServerProcess. In the docker image that class lives in
lib/tika-server-core-4.0.0.jar, so the check probes
/opt/tika-server/lib/plugins, which does not exist. The actual plugins sit next
to the standard jar in /opt/tika-server/plugins.
2. It then falls back to plugins relative to the CWD. In the image that only
works because WORKDIR /opt/tika-server happens to match.
3. If neither exists, the literal relative string plugins is passed on, and the
forked PipesServer resolves it against its own CWD, ending with an empty plugin
registry and the error above.
The javadoc says "The plugin-roots will default to a 'plugins' directory at the
same level as the server jar", which is what one would expect, but step 1
checks the level of the wrong jar. There is also no CLI or environment override
in 4.0.0, only the plugin-roots config key, so the image cannot be fixed from
the outside without shipping a config file.
Suggested fix: also probe the parent of the code-source directory (covers the
lib/ layout), or resolve against the jar named on the command line rather than
the CodeSource, and resolve the final result to an absolute path before handing
it to the pipes fork so the fork's CWD stops mattering. Independently, the
docker image could pin plugin-roots absolutely instead of relying on its
WORKDIR.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
