Description
We've observed an issue where parsing large CSV files for some reason takes a ridiculous amount of heap space and memory, seemingly unbounded (we haven't been able to set a threshold that they succeed under and we've gone up to 4gb heap space). This repros when we utilize Tika server's new JSON extraction endpoint, and it repros when we use the default TextAndCSVParser as well as if we configure it to use the older TXTParser instead. For some reason it doesn't repro when using a non-JSON extraction endpoint (though the request does still take a few minutes in that case), so I wonder if there is some recursion issue going on (didn't try with rmeta). In both cases it seems like the large file is being held as multiple character arrays in memory at once for some reason, and then there is also an extremely large object array that contains each character as a string and then some.
I have a test file that reproduces the issue, but it looks like Jira won't let me upload it (it is just under the 60mb limit but I get the "An internal error has occurred." message). I also have sample repro heap dumps that I can share (one with each parser setup) but they are definitely too large to upload to Jira at all (since they are each approximately 4gb). Let me know if there is a way I can easily share these to help showcase the issue.
Attachments
Attachments
Issue Links
- is related to
-
TIKA-3456 LanguageDetector should try to respect hasEnoughText more intelligently
- Resolved