* perf(rest): fix range download to use real seek
FileStreamingUtil.copyStreams called IOUtils.skipFully(from, start),
which reads start bytes through a 2KB scratch buffer. Combined with
ArtifactStream not overriding skip(long), a Range request at offset
600MB on an 800MB artifact made the server read+discard 600MB before
serving any payload. With 80 concurrent devices this saturated CPU.
Fix:
- ArtifactStream.skip(long) now delegates to the wrapped stream so a
FileInputStream can lseek(2). Non-seekable backends (CipherInputStream
for encrypted artifacts, S3 streams) keep their existing behaviour.
- FileStreamingUtil.copyStreams uses InputStream.skipNBytes(start)
instead of IOUtils.skipFully so the call chain reaches the underlying
skip().
JMH (single thread, 600MB offset, 1MB read): 27.21 ms -> 0.034 ms (800x).
Real stack (80 parallel curl, 1MB range at 600MB offset): avg 728 ms ->
28 ms (26x), p99 966 ms -> 54 ms.
Adds JMH test-scope dep and FileStreamingBenchmark/BufferSizeBenchmark
for regression detection. Both gated on -Dperf=true so default test
runs stay fast.
* perf(rest): drop JMH benchmarks per upstream review
Eclipse hawkBit minimizes dependencies. Drop jmh-core / jmh-generator-annprocess
test-scope deps (also GPL-2.0 — not EPL-2.0 compatible) and the two
JMH benchmarks added with the seek fix. Move the BUFFER_SIZE rationale
into an inline comment in FileStreamingUtil so the empirical reasoning
behind keeping the 8 KiB constant stays discoverable.
The benchmarks may be reintroduced as a separate PR if upstream wants a
perf-regression harness later.
663737396c