In production environments, we observed cases where the S3 client would repeatedly fail to connect due to DNS entries becoming stale. Because the existing logic only attempted the first resolved address and lacked a way to refresh DNS state, the client could get stuck in a failure loop.
Introduce RR TTL and connection failure retry to
- re-resolve the RR in a timely manner
- forcefully reset and re-resolve addresses
- add a special case when the TTL is 0 and the record must be resolved for every request
Fixes: CUSTOMER-96
Fixes: CUSTOMER-139
Should be backported to 2025.3/4 and 2026.1 since we already encountered it in the production clusters for 2025.3
- (cherry picked from commit bd9d5ad75b)
- (cherry picked from commit 359d0b7a3e)
- (cherry picked from commit ce0c7b5896)
- (cherry picked from commit 5b3e513cba)
- (cherry picked from commit 66a33619da)
- (cherry picked from commit 6eb7dba352)
- (cherry picked from commit a05a4593a6)
- (cherry picked from commit 3a31380b2c)
- (cherry picked from commit 912c48a806)
Parent PR: #27891Closesscylladb/scylladb#28403
* github.com:scylladb/scylladb:
connection_factory: includes cleanup
dns_connection_factory: refine the move constructor
connection_factory: retry on failure
connection_factory: introduce TTL timer
connection_factory: get rid of shared_future in dns_connection_factory
connection_factory: extract connection logic into a member
connection_factory: remove unnecessary `else`
connection_factory: use all resolved DNS addresses
s3_test: remove client double-close
We currently make the local node the only token owner (that owns the
whole ring) in maintenance mode, but we don't update the topology properly.
The node is present in the topology, but in the `none` state. That's how
it's inserted by `tm.get_topology().set_host_id_cfg(host_id);` in
`scylla_main`. As a result, the node started in maintenance mode crashes
in the following way in the presence of a vnodes-based keyspace with the
NetworkTopologyStrategy:
```
scylla: locator/network_topology_strategy.cc:207:
locator::natural_endpoints_tracker::natural_endpoints_tracker(
const token_metadata &, const network_topology_strategy::dc_rep_factor_map &):
Assertion `!_token_owners.empty() && !_racks.empty()' failed.
```
Both `_token_owners` and `_racks` are empty. The reason is that
`_tm.get_datacenter_token_owners()` and
`_tm.get_datacenter_racks_token_owners()` called above filter out nodes
in the `none` state.
This bug basically made maintenance mode unusable in customer clusters.
We fix it by changing the node state to `normal`.
We also extend `test_maintenance_mode` to provide a reproducer for
Fixes#27988
This PR must be backported to all branches, as maintenance mode is
currently unusable everywhere.
- (cherry picked from commit a08c53ae4b)
- (cherry picked from commit 9d4a5ade08)
- (cherry picked from commit c92962ca45)
- (cherry picked from commit 408c6ea3ee)
- (cherry picked from commit 53f58b85b7)
- (cherry picked from commit 867a1ca346)
- (cherry picked from commit 6c547e1692)
- (cherry picked from commit 7e7b9977c5)
Parent PR: #28322Closesscylladb/scylladb#28497
* https://github.com/scylladb/scylladb:
test: test_maintenance_mode: enable maintenance mode properly
test: test_maintenance_mode: shutdown cluster connections
test: test_maintenance_mode: run with different keyspace options
test: test_maintenance_mode: check that group0 is disabled by creating a keyspace
test: test_maintenance_mode: get rid of the conditional skip
test: test_maintenance_mode: remove the redundant value from the query result
storage_proxy: skip validate_read_replica in maintenance mode
storage_service: set up topology properly in maintenance mode
Clean up the awkward move constructor that was declared in the header
but defaulted in a separate compilation unit, improving clarity and
consistency.
(cherry picked from commit 3a31380b2c)
If connecting to a provided address throws, renew the address list and
retry once (and only once) before giving up.
(cherry picked from commit a05a4593a6)
Add a TTL-based timer to connection_factory to automatically refresh
resolved host name addresses when they expire.
(cherry picked from commit 6eb7dba352)
Move state management from dns_connection_factory into state class
itself to encapsulate its internal state and stop managing it from the
`dns_connection_factory`
(cherry picked from commit 66a33619da)
`test_chunked_download_data_source_with_delays` was calling `close()` on a client twice, remove the unnecessary call
(cherry picked from commit bd9d5ad75b)
The same issue as the one fixed in
394207fd69.
This one didn't cause real problems, but it's still cleaner to fix it.
(cherry picked from commit 7e7b9977c5)
We extend the test to provide a reproducer for #27988 and to avoid
similar bugs in the future.
The test slows down from ~14s to ~19s on my local machine in dev
mode. It seems reasonable.
(cherry picked from commit 867a1ca346)
In the following commit, we make the rest run with multiple keyspaces,
and the old check becomes inconvenient. We also move it below to the
part of the code that won't be executed for each keyspace.
Additionally, we check if the error message is as expected.
(cherry picked from commit 53f58b85b7)
This skip has already caused trouble.
After 0668c642a2, the skip was always hit, and
the test was silently doing nothing. This made us miss #26816 for a long
time. The test was fixed in 222eab45f8, but we
should get rid of the skip anyway.
We increase the number of writes from 256 to 1000 to make the chance of not
finding the key on server A even lower. If that still happens, it must be
due to a bug, so we fail the test. We also make the test insert rows until
server A is a replica of one row. The expected number of inserted rows is
a small constant, so it should, in theory, make the test faster and cleaner
(we need one row on server A, so we insert exactly one such row).
It's possible to make the test fully deterministic, by e.g., hardcoding
the key and tokens of all nodes via `initial_token`, but I'm afraid it would
make the test "too deterministic" and could hide a bug.
(cherry picked from commit 408c6ea3ee)
In maintenance mode, the local node adds only itself to the topology. However,
the effective replication map of a keyspace with tablets enabled contains all
tablet replicas. It gets them from the tablets map, not the topology. Hence,
`network_topology_strategy::sanity_check_read_replicas` hits
```
throw std::runtime_error(format("Requested location for node {} not in topology. backtrace {}", id, lazy_backtrace()));
```
for tablet replicas other than the local node.
As a result, all requests to a keyspace with tablets enabled and RF > 1 fail
in debug mode (`validate_read_replica` does nothing in other modes). We don't
want to skip maintenance mode tests in debug mode, so we skip the check in
maintenance mode.
We move the `is_debug_build()` check because:
- `validate_read_replicas` is a static function with no access to the config,
- we want the `!_db.local().get_config().maintenance_mode()` check to be
dropped by the compiler in non-debug builds.
We also suppress `-Wunneeded-internal-declaration` with `[[maybe_unused]]`.
(cherry picked from commit 9d4a5ade08)
We currently make the local node the only token owner (that owns the
whole ring) in maintenance mode, but we don't update the topology properly.
The node is present in the topology, but in the `none` state. That's how
it's inserted by `tm.get_topology().set_host_id_cfg(host_id);` in
`scylla_main`. As a result, the node started in maintenance mode crashes
in the following way in the presence of a vnodes-based keyspace with the
NetworkTopologyStrategy:
```
scylla: locator/network_topology_strategy.cc:207:
locator::natural_endpoints_tracker::natural_endpoints_tracker(
const token_metadata &, const network_topology_strategy::dc_rep_factor_map &):
Assertion `!_token_owners.empty() && !_racks.empty()' failed.
```
Both `_token_owners` and `_racks` are empty. The reason is that
`_tm.get_datacenter_token_owners()` and
`_tm.get_datacenter_racks_token_owners()` called above filter out nodes
in the `none` state.
This bug basically made maintenance mode unusable in customer clusters.
We fix it by changing the node state to `normal`. We also update its
rack, datacenter, and shards count. Rack and datacenter are present in the
topology somehow, but there is nothing wrong with updating them again.
The shard count is also missing, so we better update it to avoid other
issues.
Fixes#27988
(cherry picked from commit a08c53ae4b)
When the topology coordinator refreshes load_stats, it caches load_stats for every node. In case the node becomes unresponsive, and fresh load_stats can not be read from the node, the cached version of load_stats will be used. This is to allow the load balancer to have at least some information about the table sizes and disk capacities of the host.
During load_stats refresh, we aggregate the table sizes from all the nodes. This procedure calls db.find_column_family() for each table_id found in load_stats. This function will throw if the table is not found. This will cause load_stats refresh to fail.
It is also possible for a table to have been dropped between the time load_stats has been prepared on the host, and the time it is processed on the topology coordinator. This would also cause an exception in the refresh procedure.
This fixes this problem by checking if the table still exists.
Fixes: #28359
- (cherry picked from commit 71be10b8d6)
- (cherry picked from commit 92dbde54a5)
Parent PR: #28440Closesscylladb/scylladb#28469
* github.com:scylladb/scylladb:
test: add test and reproducer for load_stats refresh exception
load_stats: handle dropped tables when refreshing load_stats
This patch adds a test and reproducer for the issue where the load_stats
refresh procedure throws exceptions if any of the tables have been
dropped since load_stats was produced.
(cherry picked from commit 92dbde54a5)
When the topology coordinator refreshes load_stats, it caches load_stats
for every node. In case the node becomes unresponsive, and fresh
load_stats can not be read from the node, the cached version of
load_stats will be used. This is to allow the load balancer to
have at least some information about the table sizes and disk capacities
of the host.
During load_stats refresh, we aggregate the table sizes from all the
nodes. This procedure calls db.find_column_family() for each table_id
found in load_stats. This function will throw if the table is not found.
This will cause load_stats refresh to fail.
It is also possible for a table to have been dropped between the time
load_stats has been prepared on the host, and the time it is processed
on the topology coordinator. This would also cause an exception in the
refresh procedure.
This patch fixes this problem by checking if the table still exists.
(cherry picked from commit 71be10b8d6)
…played
Currently, if flushing hints falls within the repair cache timeout, then the flush_time is set to batchlog_manager::_last_replay. _last_replay is updated on each replay, even if some batches weren't replayed. Due to that, we risk the data resurrection.
Update _last_replay only if all batches were replayed.
Fixes: https://github.com/scylladb/scylladb/issues/24415.
Needs backport to all live versions.
- (cherry picked from commit 4d0de1126f)
- (cherry picked from commit e3dcb7e827)
Parent PR: #26793Closesscylladb/scylladb#27092
* github.com:scylladb/scylladb:
test: extend test_batchlog_replay_failure_during_repair
db: batchlog_manager: update _last_replay only if all batches were replayed
The API contract in partition_version.hh states that when dealing with
evictable entries, a real cache tracker pointer has to be passed to all
methods that ask for it. The nonpopulating reader violates this, passing
a nullptr to the snapshot. This was observed to cause a crash when a
concurrent cache read accessed the snapshot with the null tracker.
A reproducer is included which fails before and passes after the fix.
Fixes: #26847Closesscylladb/scylladb#28163
(cherry picked from commit a53f989d2f)
Closesscylladb/scylladb#28278
The test is currently flaky. It tries to get the host ID of the bootstrapping
node via the REST API after the node crashes. This can obviously fail. The
test usually doesn't fail, though, as it relies on the host ID being saved
in `ScyllaServer._host_id` at this point by `ScyllaServer.try_get_host_id()`
repeatedly called in `ScyllaServer.start()`. However, with a very fast crash
and unlucky timings, no such call may succeed.
We deflake the test by getting the host ID before the crash. Note that at this
point, the bootstrapping node must be serving the REST API requests because
`await log.wait_for("finished do_send_ack2_msg")` above guarantees that the
node has started the gossip shadow round, which happens after starting the REST
API.
Fixes#28385Closesscylladb/scylladb#28388
(cherry picked from commit a2c1569e04)
Closesscylladb/scylladb#28415
A replaced node may have pending operation on it. The replace operation
will move the node into the 'left' state and the request will never be
completed. More over the code does not expect left node to have a
request. It will try to process the request and will crash because the
node for the request will not be found.
The patch checks is the replaced node has peening request and completes
it with failure. It also changes topology loading code to skip requests
for nodes that are in a left state. This is not strictly needed, but
makes the code more robust.
Fixes#27990Closesscylladb/scylladb#28009
(cherry picked from commit bee5f63cb6)
Closesscylladb/scylladb#28177
The test is currently flaky. It tries to get the host ID of the bootstrapping
node via the REST API after the node crashes. This can obviously fail. The
test usually doesn't fail, though, as it relies on the host ID being saved
in `ScyllaServer._host_id` at this point by `ScyllaServer.try_get_host_id()`
repeatedly called in `ScyllaServer.start()`. However, with a very fast crash
and unlucky timings, no such call may succeed.
We deflake the test by getting the host ID before the crash. Note that at this
point, the bootstrapping node must be serving the REST API requests because
`await coordinator_log.wait_for("delay_node_bootstrap: waiting for message")`
above guarantees that the node has submitted the join topology request, which
happens after starting the REST API.
Fixes#28227Closesscylladb/scylladb#28233
(cherry picked from commit e503340efc)
Closesscylladb/scylladb#28309
The test is currently flaky. It incorrectly assumes that a read with
CL=LOCAL_ONE will see the data inserted by a preceding write with
CL=LOCAL_ONE in the same datacenter with RF=2.
The same issue has already been fixed for CL=ONE in
21edec1ace. The difference is that
for CL=LOCAL_ONE, only dc1 is problematic, as dc2 has RF=1.
We fix the issue for CL=LOCAL_ONE by skipping the check for dc1.
Fixes#28253
The fix addresses CI flakiness and only changes the test, so it
should be backported.
Closesscylladb/scylladb#28274
(cherry picked from commit 1f0f694c9e)
Closesscylladb/scylladb#28303
Modify test_batchlog_replay_failure_during_repair to also check
that there isn't data resurrection if flushing hints falls within
the repair cache timeout.
(cherry picked from commit e3dcb7e827)
Currently, if flushing hints falls within the repair cache timeout,
then the flush_time is set to batchlog_manager::_last_replay.
_last_replay is updated on each replay, even if some batches weren't
replayed. Due to that, we risk the data resurrection.
Update _last_replay only if all batches were replayed.
Fixes: https://github.com/scylladb/scylladb/issues/24415.
(cherry picked from commit 4d0de1126f)
reader_permit::release_base_resources() is a soft evict for the permit:
it releases the resources aquired during admission. This is used in
cases where a single process owns multiple permits, creating a risk for
deadlock, like it is the case for repair. In this case,
release_base_resources() acts as a manual eviction mechanism to prevent
permits blockings each other from admission.
Recently we found a bad interaction between release_base_resources() and
permit eviction. Repair uses both mechanism: it marks its permits as
inactive and later it also uses release_base_resources(). This partice
might be worth reconsidering, but the fact remains that there is a bug
in the reader permit which causes the base resources to be released
twice when release_base_resources() is called on an already evicted
permit. This is incorrect and is fixed in this patch.
Improve release_base_resources():
* make _base_resources const
* move signal call into the if (_base_resources_consumed()) { }
* use reader_permit::impl::signal() instead of
reader_concurrency_semaphore::signal()
* all places where base resources are released now call
release_base_resources()
A reproducer unit test is added, which fails before and passes after the
fix.
Fixes: #28083Closesscylladb/scylladb#28155
(cherry picked from commit b7bc48e7b7)
Closesscylladb/scylladb#28242
In storage_service::raft_topology_cmd_handler we pass a lambda
wrapped in coroutine::lambda to a function that creates streaming_task_impl.
The lambda is kept in streaming_task_impl that invokes it in its run
method.
The lambda captures may be destroyed before the lambda is called, leading
to use after free.
Do not wrap a lambda passed to streaming_task_impl into coroutine::lambda.
Use this auto dissociate the lambda lifetime from the calling statement.
Fixes: https://github.com/scylladb/scylladb/issues/28200.
Closesscylladb/scylladb#28201
(cherry picked from commit 65cba0c3e7)
Closesscylladb/scylladb#28241
The loop that unwraps nested exception, rethrows nested exception and saves pointer to the temporary std::exception& inner on stack, then continues. This pointer is, thus, pointing to a released temporary
Closesscylladb/scylladb#28143
(cherry picked from commit 829bd9b598)
Closesscylladb/scylladb#28239
It is observed that:
repair - repair[667d4a59-63fb-4ca6-8feb-98da49946d8b]: Failed to update
system.repair_history table of node d27de212-6f32-4649ad76-a9ef1165fdcb:
seastar::rpc::remote_verb_error
(repair[667d4a59-63fb-4ca6-8feb-98da49946d8b]: range (minimum
token,maximum token) is not in the format of (start, end])
This is because repair checks the end of the range to be repaired needs
to be inclusive. When small_table_optimization is enabled for regular
repair, a (minimum token,maximum token) will be used.
To fix, we can relax the check of (start, end] for the min max range.
Fixes#27220
Backport to all active branches.
(cherry picked from commit e97a504)
Parent PR: #27357Closesscylladb/scylladb#27460
The `make_key` lambda erroneously allocates a fixed 8-byte buffer
(`sizeof(s.size())`) for variable-length strings, potentially causing
uninitialized bytes to be included. If such bytes exist and they are
not valid UTF-8 characters, deserialization fails:
```
ERROR 2026-01-16 08:18:26,062 [shard 0:main] testlog - snapshot_list_contains_dropped_tables: cql env callback failed, error: exceptions::invalid_request_exception (Exception while binding column p1: marshaling error: Validation failed - non-UTF8 character in a UTF8 string, at byte offset 7)
```
Fixes#28195.
Signed-off-by: Nikos Dragazis <nikolaos.dragazis@scylladb.com>
Closesscylladb/scylladb#28197
(cherry picked from commit 8aca7b0eb9)
Closesscylladb/scylladb#28208
Consider the following scenario:
1. A table has RF=3 and writes use CL=QUORUM
2. One node is down
3. There is a pending tablet migration from the unavailable node
that is reverted
During the revert, there can be a time window where the pending replica
being cleaned up still accepts writes. This leads to write failures,
as only two nodes (out of four) are able to acknowledge writes.
This patch fixes the issue by adding a barrier to the cleanup_target
tablet transition state, ensuring that the coordinator switches back to
the previous replica set before cleanup is triggered.
Fixes https://github.com/scylladb/scylladb/issues/26512
It's a pre existing issue. Backport is required to all recent 2025.x versions.
- (cherry picked from commit 669286b1d6)
- (cherry picked from commit 67f1c6d36c)
- (cherry picked from commit 6163fedd2e)
Parent PR: #27413Closesscylladb/scylladb#27427
* github.com:scylladb/scylladb:
topology_coordinator: Fix the indentation for the cleanup_target case
topology_coordinator: Add barrier to cleanup_target
test_node_failure_during_tablet_migration: Increase RF from 2 to 3
Fixes#27992
When doing a commit log oversized allocation, we lock out all other writers by grabbing
the _request_controller semaphore fully (max capacity).
We thereafter assert that the semaphore is in fact zero. However, due to how things work
with the bookkeep here, the semaphore can in fact become negative (some paths will not
actually wait for the semaphore, because this could deadlock).
Thus, if, after we grab the semaphore and execution actually returns to us (task schedule),
new_buffer via segment::allocate is called (due to a non-fully-full segment), we might
in fact grab the segment overhead from zero, resulting in a negative semaphore.
The same problem applies later when we try to sanity check the return of our permits.
Fix is trivial, just accept less-than-zero values, and take same possible ltz-value
into account in exit check (returning units)
Added whitebox (special callback interface for sync) unit test that provokes/creates
the race condition explicitly (and reliably).
Closesscylladb/scylladb#27998
(cherry picked from commit a7cdb602e1)
Closesscylladb/scylladb#28097
Currently, tablet allocation intentionally ignores current load (
introduced by the commit #1e407ab) which could cause identical shard
selection when allocating a small number of tablets in the same topology.
When a tablet allocator is asked to allocate N tablets (where N is smaller
than the number of shards on a node), it selects the first N lowest shards.
If multiple such tables are created, each allocator run picks the same
shards, leading to tablet imbalance across shards.
This change initializes the load sketch with the current shard load,
scaled into the [0,1] range, ensuring allocation still remains even
while starting from globally least-loaded shards.
Fixes https://github.com/scylladb/scylladb/issues/27620
Closes https://github.com/scylladb/scylladb/pull/27802Closesscylladb/scylladb#28106
`test_schema_versioning_with_recovery` is currently flaky. It performs
a write with CL=ALL and then checks if the schema version is the same on
all nodes by calling `verify_table_versions_synced`. All nodes are expected
to sync their schema before handling the replica write. The node in
RECOVERY mode should do it through a schema pull, and other nodes should do
it through a group 0 read barrier.
The problem is in `verify_local_schema_versions_synced` that compares the
schema versions in `system.local`. The node in RECOVERY mode updates the
schema version in `system.local` after it acknowledges the replica write
as completed. Hence, the check can fail.
We fix the problem by making the function wait until the schema versions
match.
Note that RECOVERY mode is about to be retired together with the whole
gossip-based topology in 2026.2. So, this test is about to be deleted.
However, we still want to fix it, so that it doesn't bother us in older
branches.
Fixes#23803Closesscylladb/scylladb#28114
(cherry picked from commit 6b5923c64e)
Closesscylladb/scylladb#28175
We currently do it only for a bootstrapping node, which is a bug. The
missing IP can cause an internal error, for example, in the following
scenario:
- replace fails during streaming,
- all live nodes are shut down before the rollback of replace completes,
- all live nodes are restarted,
- live nodes start hitting internal error in all operations that
require IP of the replacing node (like client requests or REST API
requests coming from nodetool).
We fix the bug here, but we do it separately for replace with different
IP and replace with the same IP.
For replace with different IP, we persist the IP -> host ID mapping
in `system.peers` just like for bootstrap. That's necessary, since there
is no other way to determine IP of the replacing node on restart.
For replace with the same IP, we can't do the same. This would require
deleting the row corresponding to the node being replaced from
`system.peers`. That's fine in theory, as that node is permanently
banned, so its IP shouldn't be needed. Unfortunately, we have many
places in the code where we assume that IP of a topology member is always
present in the address map or that a topology member is always present in
the gossiper endpoint set. Examples of such places:
- nodetool operations,
- REST API endpoints,
- `db::hints::manager::store_hint`,
- `group0_voter_handler::update_nodes`.
We could fix all those places and verify that drivers work properly when
they see a node in the token metadata, but not in `system.peers`.
However, that would be too risky to backport.
We take a different approach. We recover IP of the replacing node on
restart based on the state of the topology state machine and
`system.peers` just after loading `system.peers`.
We rely on the fact that group 0 is set up at this point. The only case
where this assumption is incorrect is a restart in the Raft-based
recovery procedure. However, hitting this problem then seems improbable,
and even if it happens, we can restart the node again after ensuring
that no client and REST API requests come before replace is rolled back
on the new topology coordinator. Hence, it's not worth to complicate the
fix (by e.g. looking at the persistent topology state instead of the
in-memory state machine).
Fixes#28057
Backport this PR to all branches as it fixes a problematic bug.
- (cherry picked from commit fc4c2df2ce)
- (cherry picked from commit 4526dd93b1)
- (cherry picked from commit 749b0278e5)
- (cherry picked from commit 0fed9f94f8)
Manually cherry-picked:
- 90b5b2c5f5
- 92b165b8c0
Parent PR: #27435Closesscylladb/scylladb#28098
* https://github.com/scylladb/scylladb:
gossiper: add_saved_endpoint: make generations of excluded nodes negative
test: introduce test_full_shutdown_during_replace
utils: error_injection: allow aborting wait_for_message
raft topology: preserve IP -> ID mapping of a replacing node on restart
pylib/rest_client.py: encode injection name
utils/error_injection: allow to abort `injection_handler::wait_for_message()`
