⟳ Updates Hub
S
Home / Database

ScyllaDB

Website ↗ RSS 80 items Fetched 2h ago
S
ScyllaDB Database v2026.2.10

scylla-2026.2.10

Merge '[Backport 2026.2] alternator: report WarmThroughput and GSI Pr…

S
ScyllaDB Database v2026.1.15

scylla-2026.1.15

service: tablet_allocator: check a migration group with its combined …

S
ScyllaDB Database v2026.3.3

scylla-2026.3.3: commitlog: Make replay large fragment cache per-(segment-)shard

Fixes: SCYLLADB-4945 Fragment ID for large commitlog entries are only unique per-shard. For a non-corrupt commitlog, this is not a problem, since we replay in descriptor order -> partitioned per shard. But if one or more segments are broken, we could end up with lingering data from one shard sharing id with a subsequent one from another shard. This can cause data corruption or data loss, depending…

S
ScyllaDB Database v2026.3.3-candidate

scylla-2026.3.3-candidate-20261004023307: commitlog: Make replay large fragment cache per-(segment-)shard

Fixes: SCYLLADB-4945 Fragment ID for large commitlog entries are only unique per-shard. For a non-corrupt commitlog, this is not a problem, since we replay in descriptor order -> partitioned per shard. But if one or more segments are broken, we could end up with lingering data from one shard sharing id with a subsequent one from another shard. This can cause data corruption or data loss, depending…

S
ScyllaDB Database v2025.1.16

scylla-2025.1.16: topology_coordinator: don't run ordered ops under a foreign session

handle_topology_ordered_op() releases the group0 guard before sending its RPCs, but reads _topology.session afterwards. That read is covered neither by the guard's read_apply mutex nor by the term check in start_operation(), so a coordinator which lost leadership in that window can pick up the session installed by a later, unrelated operation. Replicas accept the RPC because the session exists and…

S
ScyllaDB Database v2025.1.16-candidate

scylla-2025.1.16-candidate-20260925095216: topology_coordinator: don't run ordered ops under a foreign session

handle_topology_ordered_op() releases the group0 guard before sending its RPCs, but reads _topology.session afterwards. That read is covered neither by the guard's read_apply mutex nor by the term check in start_operation(), so a coordinator which lost leadership in that window can pick up the session installed by a later, unrelated operation. Replicas accept the RPC because the session exists and…

S
ScyllaDB Database v2026.1.14

scylla-2026.1.14: commitlog: Handle abandoned future in new_segment

Fixes: SCYLLADB-4525 The timeout handling in active_segmentwill, on fail abandon a pending new_segment call, which will (probably) still finish eventually. However, if the odds play against us, we could end up with it finishing when another segment has already been added as active. This is particularly true on large allocation, where we restore segment state and can thusly end up leaving an alloca…

S
ScyllaDB Database v2026.1.14-candidate

scylla-2026.1.14-candidate-20260925095214: commitlog: Handle abandoned future in new_segment

Fixes: SCYLLADB-4525 The timeout handling in active_segmentwill, on fail abandon a pending new_segment call, which will (probably) still finish eventually. However, if the odds play against us, we could end up with it finishing when another segment has already been added as active. This is particularly true on large allocation, where we restore segment state and can thusly end up leaving an alloca…

S
ScyllaDB Database v2026.2.8

scylla-2026.2.8

Merge '[Backport 2026.2] alternator: keep a retained CDC log out of L…

S
ScyllaDB Database v2026.3.2

scylla-2026.3.2

Merge '[Backport 2026.3] cql3: validate compaction sub-options for vi…

S
ScyllaDB Database v2026.3.1

scylla-2026.3.1: test/cluster: accept the ban notification when a join is rolled back

A node whose bootstrap or replace fails is rolled back by the topology coordinator through left_token_ring, and finish_left_token_ring_transition removes it from group0 and bans it. The joining node is racing that: it either observes its own failed request and aborts startup with the message the test expects, or receives the ban notification first and terminates on that instead. Both are legitimat…

S
ScyllaDB Database v2026.3.1-candidate

scylla-2026.3.1-candidate-20260906035304: test/cluster: accept the ban notification when a join is rolled back

A node whose bootstrap or replace fails is rolled back by the topology coordinator through left_token_ring, and finish_left_token_ring_transition removes it from group0 and bans it. The joining node is racing that: it either observes its own failed request and aborts startup with the message the test expects, or receives the ban notification first and terminates on that instead. Both are legitimat…