Dawid Weiss created COMPRESS-623:
------------------------------------
Summary: Speed up ZipFile's raw stream copying by not seeking for
local headers until they're needed
Key: COMPRESS-623
URL: https://issues.apache.org/jira/browse/COMPRESS-623
Project: Commons Compress
Issue Type: Improvement
Reporter: Dawid Weiss
I have a somewhat odd use case with gigabytes of ZIP files, each with thousands
of documents (on comparatively slow, network drives). We need to restructure
these ZIPs without the need to recompress files.
The above turns out to work almost perfectly with raw-data copying ZipFile
offers but empirical tests showed a major slowdown in the initial opening of
zip files, linked to multiple reads/seeks for local file headers. If an option
is passed to ignore those headers, raw streams are inaccessible.
I've taken a look at the code and the code in getRawInputStream could basically
do the same thing that getInputStream does - lazily load the missing offset via
getDataOffset(ZipEntry). In fact, getInputStream could just call
getRawInputStream directly, which avoids some code duplication.
I also restructured the code that maintains LinkedList for duplicated names a
bit. I think this bit could be removed altogether and replaced with a
pointer-chain list directly on Entry class instances (the map would then point
at the first entry; duplicated entries are an oddity and infrequent in the
wild).
I see speedups for opening and copying random raw streams in the order of 3-4x
and all the current tests pass. I filed a PR at github - happy to discuss it
there.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)