Pavel Emelyanov

Recently, the sink impls were fixed not to crash in case caller tries to use zero-copy writing (52eb05829b271984dd4c79dae940057e70c13c2c), but the chunked and content-length sinks both use fallback implementation of `put(net::packet)`...

Lrge allocation threshold never goes down

1

It's designed like that, but it has a side effect -- rare large allocations may shadow more interesting smaller ones forever. E.g. -- if startup allocates large chunk of memory,...

posix_datagram_channel::send() is not fiber-safe

11

```c++ future posix_datagram_channel::send(const socket_address& dst, packet p) { auto len = p.len(); _send.prepare(dst, std::move(p)); auto sg_id = internal::scheduling_group_index(current_scheduling_group()); bytes_sent[sg_id] += len; return _fd.sendmsg(&_send._hdr) .then([len] (size_t size) { SEASTAR_ASSERT(size == len);...

io_uring backend is too ENOMEM-y in older kernels, leading to unwanted errors

4

There's a fix in Linux that addresses that (https://github.com/torvalds/linux/commit/182db972c9568dc530b2f586a2f82dfd039d9f2a), but for older kernels it's better to ban io-uring backend and fall-back to AIO one

It just std::move-s a buffer and a semaphore_units objects, both moves are noexcept, so is the constructor itself.

Flaky test topology_operation_during_upgrade_test.TopologyOperationWithMixedCLusterTest.test_trigger_snapshot_transfer

2

While validating seastar update (#24541) got error: ``` ccmlib.node.ToolError: Subprocess /jenkins/workspace/scylla-2025.2/gating-dtest-release-with-consistent-topology-changes/scylla/.dtest/dtest-_hzj3xeg/test/node1/bin/scylla sstable dump-data --merge /jenkins/workspace/scylla-2025.2/gating-dtest-release-with-consistent-topology-changes/scylla/.dtest/dtest-_hzj3xeg/test/node1/data/ks/cf-c6ad76104b7c11f0a0fe9c4b585881b3/me-3gr7_10rb_5mjo126b0q4vw7jtlc-big-Data.db /jenkins/workspace/scylla-2025.2/gating-dtest-release-with-consistent-topology-changes/scylla/.dtest/dtest-_hzj3xeg/test/node1/data/ks/cf-c6ad76104b7c11f0a0fe9c4b585881b3/me-3gr7_10rb_5mjo12l6x8sf9wfmhc-big-Data.db /jenkins/workspace/scylla-2025.2/gating-dtest-release-with-consistent-topology-changes/scylla/.dtest/dtest-_hzj3xeg/test/node1/data/ks/cf-c6ad76104b7c11f0a0fe9c4b585881b3/me-3gr7_10rb_5n6td26b0q4vw7jtlc-big-Data.db /jenkins/workspace/scylla-2025.2/gating-dtest-release-with-consistent-topology-changes/scylla/.dtest/dtest-_hzj3xeg/test/node1/data/ks/cf-c6ad76104b7c11f0a0fe9c4b585881b3/me-3gr7_10rb_5n6td2l6x8sf9wfmhc-big-Data.db exited with non-zero status; exit status: 1; stderr: error loading sstables: std::filesystem::__cxx11::filesystem_error...

Query latency raises during native backup

44

Created to track the research. Latencies p90/p99 in ms | | Read | Write | |-|-|-| | Baseline (pure workload) | 3.34 / 5.44 | 1.71 / 2.40 | |...

Move backup and restore workflow into own sched group

3

Nowadays backup (and restore) live in the streaming sched group thus sharing CPU and IO with e.g. repair activity as well as nodetool-triggered compaction. As [we've observed](https://github.com/scylladb/scylladb/issues/24419#issuecomment-2955515721), uploading files can...

[backport 2025.2] Update seastar submodule (no nested stall backtraces)

2

* seastar 9f0034a0...450e36d5 (1): > stall_detector: no backtrace if exception Fixes #24464

Add ability to keep backup IO throttled

16

Nowadays backup lives in the streaming sched group thus sharing CPU and IO with e.g. repair activity or nodetool-triggered compaction. It was observed, that uploading files can easily fill up...