* Rework `needle_map.CompactMap()` to maximize memory efficiency.
* Use a memory-efficient structure for `CompactMap` needle value entries.
This slightly complicates the code, but makes a **massive** difference
in memory efficiency - preliminary results show a ~30% reduction in
heap usage, with no measurable performance impact otherwise.
* Clean up type for `CompactMap` chunk IDs.
* Add a small comment description for `CompactMap()`.
* Add the old version of `CompactMap()` for comparison purposes.
Per-entry memory usage is based on `TotalAllocs`, which is incorrect - that
value is a cummulative of heap usage, which doesn't decrease when objects
are freeed.
`Allocs` is instead an accurate represeentation of actual memory usage
at the time metrics are reported.
Changed the signal from SIGUSR1 to SIGTERM. This should fix the compilation error since SIGTERM is available on all Unix-like systems including macOS. The functionality remains similar - it will still allow the master process to wait for a signal before exiting, just using a different signal type.
Improve safety for weed shells `ec.encode`.
The current process for `ec.encode` is:
1. EC shards for a volume are generated and added to a single server
2. The original volume is deleted
3. EC shards get re-balanced across the entire topology
It is then possible to lose data between #2 and #3, if the underlying volume storage/server/rack/DC
happens to fail, for whatever reason. As a fix, this MR reworks `ec.encode` so:
* Newly created EC shards are spread across all locations for the source volume.
* Source volumes are deleted only after EC shards are converted and balanced.
NeedleHeaderSize happen to have the same size as NeedleMapEntrySize, except when running the 5 bytes offset variant of Seaweedfs, because it does not contain OffsetSize. This causes ECX corruption on deletes, due to the drifting offset computation (offset is always computed on a basis of 16 bytes per record instead of 17 bytes)
Signed-off-by: Quentin Devos <4972091+Okhoshi@users.noreply.github.com>
Improve parallelization for `ec.encode`.
Instead of processing one volume at at time, perform all EC conversion
steps (mark readonly -> generate EC shards -> delete volume -> remount) in
parallel for all of them.
This should substantially improve performance when EC encoding
entire collections.
chrislu
Lisandro Pin
SmoothDenis
Chris Lu
Luna Yao
FQHSLycopene
Aleksey Kosov
Bruce Zou
royatwp
Mohamed Sekour
Weihao Jiang
bwlfhu
dongxufeng
Kuzmin Anton
Quentin D.
ARibster