process_exporter input plugin does not handle dead processes

[PromyLOPh]

Bug Report

Describe the bug
The process_explorer input plugin only adds data to its own internal state (like flb_pe.memory_bytes), but never removes data from dead processes. If running for a long time, this leads to high memory usage, because old process data is never pruned, and high CPU usage, because cmetric’s metric_hash_lookup function is not actually a hash lookup, but linear search. With a growing number of stale process entries for each metric the fluent-bit process will end up spending most of its time doing this search.

To Reproduce

• Run fluent-bit -p scrape_interval=1 -i process_exporter_metrics -o stdout
• Run while true; do for i in $(seq 100); do sleep 5 & done; sleep 5; done at the same time to create alot of short-lived processes.
• Wait about 10–15 minutes.
• Observe increasing CPU and memory usage of the fluent-bit process via htop or similar. Run ptrace record -p <pid> to capture profiling data for a few seconds, run ptrace report and note that the function metric_hash_lookup is on the top of the list.

Expected behavior
CPU and memory usage stays the same.

Your Environment

• Version used: 3.1.4 and latest master (commit 104aceb)
• Configuration: (none)
• Environment name and version (e.g. Kubernetes? What version?): (none)
• Server type and version: (none)
• Operating System and version: Debian 12
• Filters and plugins: plugin_exporter_metrics and stdout

Additional context
fluent-bit’s process_exporter plugin is basically unusable on low-end embedded devices. Also note that during testing fluent-bit might crash due to a different bug (see #9548).