Seen this issue in Karaf 4.4.9, but suspect it will impact newer releases.
SimpleDownloadTask.download() stages wrap:/blueprint:/spring: bundle URLs into a file named by hashing the URL, then does:
if (file.exists() && !file.delete()) { throw ...; }
tmpFile.renameTo(file);
If two overlapping resolutions of the same URL race (e.g. two feature installs, each with their own DownloadManager — dedup only happens within one instance), the second one can delete() the file the first one just wrote, right as something else opens it:
java.io.FileNotFoundException: /..../data/tmp/9c73e02c-failureaccess-1.0.1.jar (No such file or directory)
at java.base/java.io.FileInputStream.open0(Native Method)
at java.base/java.io.FileInputStream.open(FileInputStream.java:216)
Seen for multiple different artifacts, so it's generic to the staging code, not any one jar. Shows up as needing multiple restarts before a container boots cleanly.
Fix
Replace delete-then-rename with one atomic move:
try {
Files.move(tmpFile.toPath(), file.toPath(),
StandardCopyOption.REPLACE_EXISTING, StandardCopyOption.ATOMIC_MOVE);
} catch (AtomicMoveNotSupportedException e) {
Files.move(tmpFile.toPath(), file.toPath(), StandardCopyOption.REPLACE_EXISTING);
}
A concurrent reader then always sees either the old or new file — never a missing one.
Proof
Stress test (attached RaceDemo.java, 6 writers + 6 readers, 3s):
original delete+renameTo -> FileNotFoundException count: 18733
patched Files.move -> FileNotFoundException count: 0
Attached
Related: #2805 (different bug, same "startup race" theme — fileinstall config corruption).
Seen this issue in Karaf 4.4.9, but suspect it will impact newer releases.
SimpleDownloadTask.download()stageswrap:/blueprint:/spring:bundle URLs into a file named by hashing the URL, then does:If two overlapping resolutions of the same URL race (e.g. two feature installs, each with their own
DownloadManager— dedup only happens within one instance), the second one candelete()the file the first one just wrote, right as something else opens it:Seen for multiple different artifacts, so it's generic to the staging code, not any one jar. Shows up as needing multiple restarts before a container boots cleanly.
Fix
Replace delete-then-rename with one atomic move:
A concurrent reader then always sees either the old or new file — never a missing one.
Proof
Stress test (attached
RaceDemo.java, 6 writers + 6 readers, 3s):Attached
javac RaceDemo.java && java RaceDemo)Related: #2805 (different bug, same "startup race" theme — fileinstall config corruption).