The patch series introduces linearisable topology changes using
raft protocol. The state machine driven by raft is described in
"service: Introduce topology state machine". Some explanations about
the implementation can be found in "storage_service: raft topology:
implement topology management through raft".
The code is not ready for production. There is not much in terms of error
handling and integration with the rest of the system is not even started.
For full integration request fencing will need to be implemented and
token_metadata has to be extended to support not just "pending" nodes
but concepts of "read replica set" and "write replica set".
The code may be far from be usable, but it is hidden behind the
"experimental raft" flag and having it in tree will relieve me from
constant rebase burden.
* 'raft-topology-v6' of github.com:scylladb/scylla-dev:
storage_service: fix indentation from previous patch
storage_service: raft topology: implement topology management through raft
service: raft: make group0_guard move assignable
service: raft: wire up apply() and snapshot transfer for topology in group0 state machine
storage_service: raft topology: introduce a function that applies topology cmd to local state machine
storage_service: raft topology: introduce a raft monitor and topology coordinator fibers
storage_service: raft topology: introduce snapshot transfer code for the topology table
raft topology: add RAFT_TOPOLOGY_CMD verb that will be used by topology coordinator to communicated with nodes
bootstrapper: Add get_random_bootstrap_tokens function
service: raft: add support for topology_change command into raft_group0_client
service: raft: introduce topology_change group0 command
system_keyspace: add a table to persist topology change state machine's state
service: Introduce topology state machine data structures
storage_proxy: not consult topology on local table write
The code here implements the state machine described in "service:
Introduce topology state machine". A topology operation is requested
by writing into topology_request field through raft. After that
topology_change_transition() function running on a leader is responsible
to drive the operation to completion. There is no much in terms of error
handling here yet. It something fails the code will just continue trying.
topology_change_state_load() which is (eventually) called on all nodes each
time state machine's state changes is a glue between the raft view of
the topology and the rest of the "legacy" system. The code there creates
token_metadata object from the raft view and fills in peers table which
is needed for drivers. The gossiper is almost completely cut of from the
topology management, but the code still updates node's sate there to
'normal' and 'left' for some legacy functionality to continue working.
Note that handlers for those states are disabled in raft mode.
raft_topology_cmd_handler() is called by topology coordinator and this
is where the streaming happens. The kind of streaming depends on the
state the node is in. The function is "re-entrable". It can be called
more then once and will either start new operation if it is the first
invocation or previous one failed, or it will wait from previous
operation to complete.
The new code is hidden behind "experimental raft" and should not change
how the system works if disabled.
Some indentation here is intentionally left wrong and will be fixed by
the next patch.
The function applies to persistent storage and call stub function
topology_change_state_load() that will load the new state into the
memory in later patches.
Raft monitor fiber monitors local's server raft state and starts the
topology coordinator fiber when it becomes a leader. Stops it when it
is not longer a leader.
The coordinator fiber waits for topology state changes, but there will
be none yet.
Empty for now. Will be used later by the topology coordinator to
communicate with other nodes to instruct them to start streaming,
or start to fence read/writes.
This patch increases the connection timeout in the get_cql_cluster()
function in test/cql-pytest/run.py. This function is used to test
that Scylla came up, and also test/alternator/run uses it to set
up the authentication - which can only be done through CQL.
The Python driver has 2-second and 5-second default timeouts that should
have been more than enough for everybody (TM), but in #13239 we saw
that in one case it apparently wasn't enough. So to be extra safe,
let's increase the default connection-related timeouts to 60 seconds.
Note this change only affects the Scylla *boot* in the test/*/run
scripts, and it does not affect the actual tests - those have different
code to connect to Scylla (see cql_session() in test/cql-pytest/util.py),
and we already increased the timeouts there in #11289.
Fixes#13239
Signed-off-by: Nadav Har'El <nyh@scylladb.com>
Closes#13291
This reverts commit c6087cf3a0.
Said commit can cause a deadlock when 2 or more repairs compete for
locks on 2 or more nodes. Consider the following scenario:
Node n1 and n2 in the cluster, 1 shard per node, rf = 2, each shard has
1 available unit for the reader lock
n1 starts repair r1
r1-n1 (instance of r1 on node1) takes the reader lock on node1
n2 starts repair r2
r2-n2 (instance of r2 on node2) takes the reader lock on node2
r1-n2 will fail to take the reader lock on node2
r2-n1 will fail to take the reader lock on node1
As a result, r1 and r2 could not make progress and deadlock happens.
The complexity comes from the fact that a repair job needs lock on more
than one node. It is not guaranteed that all the participant nodes could
take the lock in one short.
There is no simple solution to this so we have to revert this locking
mechanism and look for another way to prevent reader trashing when
repairing nodes with mismatching shard count.
Fixes: #12693Closes#13266
Cassandra detects when a batch has both an IF EXISTS and IF NOT EXISTS
on the same row, and complains this is not a useful request (after all,
it can never succeed, because the batch can only succeed if both conditions
are true, and that can't be if one checks IF EXISTS and the other
IF NOT EXISTS).
This patch adds a test, test_lwt_with_batch_conflict_1, which checks
that this case results in an error. It passes on Cassandra, but xfails
on Scylla which doesn't report an error in this case.
A second test, test_lwt_with_batch_conflict_2, shows that the detection
of the EXISTS / NOT EXISTS conflict is special, and other conflicts
such as having both "r=1" and "r=2" for the same row, are NOT detected
by Cassandra.
Refs #13011.
Signed-off-by: Nadav Har'El <nyh@scylladb.com>
Closes#13270
It's declared in header, but is not used outside of .cc. Forward
declaration in header would be enough.
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
Closes#13289
for better readability.
also, add `#include <concepts>`, as we should include what we use
instead of relying on other headers do this on behalf of us.
Signed-off-by: Kefu Chai <kefu.chai@scylladb.com>
Closes#13277
clang warns when the implicit conversion changes the precision of the
converted number. in this case, the before being multiplied,
`std::numeric_limits<unsigned long>::max() >> 1` is implicitly
promoted to double so it can obtain the common type of double and
unsigned long. and the compiler warns:
```
/home/kefu/dev/scylladb/test/boost/network_topology_strategy_test.cc:129:84: error: implicit conversion from 'unsigned long' to 'double' changes value from 9223372036854775807 to 9223372036854775808 [-Werror,-Wimplicit-const-int-float-conversion]
return static_cast<unsigned long>(d*(std::numeric_limits<unsigned long>::max() >> 1)) << 1;
~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~
```
but
1. we don't really care about the precision here, we just want to map a
double to a token represented by an int64_t
2. the maximum possible number being converted is less than
9223372036854775807, which is the maximum number of int64_t, which
is in general an alias of `long long`, not to mention that
LONG_MAX is always 2147483647, after shifting right, the result
would be 1073741823
so this is a false alarm. in order to silence it, we explicitly
cast the RHS of `*` operator to double.
Signed-off-by: Kefu Chai <kefu.chai@scylladb.com>
Closes#13221
Hey y'all!
Me and @malusev998 are maintaining a updated version of the [PHP Driver ](https://github.com/he4rt/scylladb-php-driver) together with @he4rt community and it had a bunch of improvements on these last month.
Before it was working only at PHP 7.1 (DataStax branch), and at our branch we have it working at PHP 8.1 and 8.2.
We are also using the ScyllaDB C++ Driver on this project and I think that is a good idea to point new users for this project since it's the most updated PHP Driver maintained now.
What do y'all think about that?
Closes#13218
* github.com:scylladb/scylladb:
fix: links to php driver
fix: adding php versions into driver's description
docs: scylladb better php driver
this change is a leftover of 063b3be8a7, which failed to include the changes in the header files.
it turns out we have `using namespace httpd;` in seastar's `request_parser.rl`, and we should not rely on this statement to expose the symbols in `seatar::httpd` to `seastar` namespace. in this change,
also, sine `get_name()` previously a non-static member function of `seastar_test` is now a static member function, so we need to update the tests which capture `this` for calling this function, so they don't capture `this` anymore.
Closes#13202
* github.com:scylladb/scylladb:
test: drop unused captured variables
Update seastar submodule
* seastar 9cbc1fe889...1204efbc5e (14):
> http: Add lost pragma once into client.hh
> prometheus, http: do not expose httpd::* in seastar
> build: add haswell support
> ci: fix configuration to build checkheaders target.
> core: map_reduce: Fix use-after-free in variant with futurized reducer
> Merge 'tests: support boost::test decorators and tolerate failures in test_spawn_input' from Kefu Chai
> memory: support reallocing foreign (non-Seastar) memory on a reactor thread
> test: futures: disable -Wself-move for GCC>=13
> map_reduce: do not move a temporary object
> doc/building-dpdk.md: drop extraneous '$'
> http: url_decode: translate plus back into char
> Merge 'seastar-json2code: cleanups' from Kefu Chai
> Fix markdown formatting
> Merge 'Minor abort on OOM changes' from Travis Downs
This patch extends a previous patch that added these metrics globally:
- cql_requests_count
- cql_request_bytes
- cql_response_bytes
This patch adds a "scheduling_group_name" label to these metrics and changes corresponding
counters to be accounted on a per-scheduling-group level.
As a bonus this patch also marks all 3 metrics as 'skip_when_empty'.
Ref #13061
Signed-off-by: Vlad Zolotarov <vladz@scylladb.com>
Message-Id: <20230321201412.3004845-1-vladz@scylladb.com>
Notably, to admission execution and eviction. Registering/unregistering
the permit as inactive is not traced, as this happens on every
buffer-fill for range scans.
Semaphore trace messages have a "[reader_concurrency_semaphore]" prefix
to allow them to be clearly associated with the semaphore.
To make sure all tracing done on a certain page will make its way into
the appropriate trace session.
This is a contination of the previous patch (which added trace pointer
to the permit).
And propagate it down to where it is created. This will be used to add
trace points for semaphore related events, but this will come in the
next patches.
perf.cc has two key comparators: key_compare and key_tri_compare. These
are very generic name, in fact key_compare directly clashes with a
comparator with the same name in types.hh. Avoid the clash by renaming
both of these to a more unique name.
This is a translation of Cassandra's CQL unit test source file
validation/operations/SelectMultiColumnRelationTest.java into our
cql-pytest framework.
The tests reproduce four already-known Scylla bugs and three new bugs.
All tests pass on Cassandra. Because of these bugs 9 of the 22 tests
are marked xfail, and one is marked skip (it crashes Scylla).
Already known issues:
Refs #64: CQL Multi column restrictions are allowed only on a clustering
key prefix
Refs #4178: Not covered corner case for key prefix optimization in filtering
Refs #4244: Add support for mixing token, multi- and single-column
restrictions
Refs #8627: Cleanly reject updates with indexed values where value > 64k
New issue discovered by these tests:
Refs #13217: Internal server error when null is used in multi-column relation
Refs #13241: Multi-column IN restriction with tuples of different lengths
crashes Scylla
Refs #13250: One-element multi-column restriction should be handled like a
single-column restriction
Signed-off-by: Nadav Har'El <nyh@scylladb.com>
Closes#13265
Point to the difference between the official MurmurHash3 and Scylla / Cassandra implementation
Update docs/glossary.rst
Co-authored-by: Anna Stuchlik <37244380+annastuchlik@users.noreply.github.com>
Closes#11369
Call replicate_live_endpoints on shard 0 to copy from 0 to the rest of
the shards. And get the list of live members from shard 0.
Move lock to the callers.
Signed-off-by: Alejo Sanchez <alejo.sanchez@scylladb.com>
Closes#13240
The topology state machine will track all the nodes in a cluster,
their state, properties (topology, tokens, etc) and requested actions.
Node state can be one of those:
none - the node is not yet in the cluster
bootstrapping - the node is currently bootstrapping
decommissioning - the node is being decommissioned
removing - the node is being removed
replacing - the node is replacing another node
normal - the node is working normally
rebuild - the node is being rebuilt
left - the node is left the cluster
Nodes in state left are never removed from the state.
Tokens also can be in one of the states:
write_both_read_old - writes are going to new and old replica, but reads are from
old replicas still
write_both_read_new - writes still going to old and new replicas but reads are
from new replica
owner - tokens are owned by the node and reads and write go to new
replica set only
Tokens that needs to be move start in 'write_both_read_old' state. After entire
cluster learns about it streaming start. After the streaming tokens move
to 'write_both_read_new' state and again the whole cluster needs to learn about it
and make sure no reads started before that point exist in the system.
After that tokens may move to 'owner' state.
topology_request is the field through which a topology operation request
can be issued to a node. A request is one of the topology operation
currently supported: join, leave, replace or remove.
Writes to tables with local replication strategies do not need to consult
the topology. This is not only an optimization but it allows writing
into the local tables before topology is known.
this change is a leftover of 063b3be,
which failed to include the changes in the header files.
it turns out we have `using namespace httpd;` in seastar's
`request_parser.rl`, and we should not rely on this statement to
expose the symbols in `seatar::httpd` to `seastar` namespace.
in this change,
* api/*.hh: all httpd symbols are referenced by `httpd::*`
instead of being referenced as if they are in `seastar`.
Signed-off-by: Kefu Chai <kefu.chai@scylladb.com>
There's a need to convert both -- version and format -- to string and back. Currently, there's a disperse set of helpers in sstables/ code doing that and this PR brings some other to it
- adds fmt::formatter<> specialization for both types
- leaves one set of {format|version}_from_string() helpers converting any string-ish object into value
refs: #12523Closes#13214
* github.com:scylladb/scylladb:
sstables: Expell sstable_version_types from_string() helper
sstables: Generalize ..._from_string helpers
sstables: Implement fmt::formatter<sstable_format_types>
sstables: Implement fmt::formatter<sstable_version_types>
sstables: Move format maps to namespace scope
In the future, when testing WASM UDFs, we will only store the Rust
source codes of them, and compile them to WASM. To be able to
do that, we need rust standard library for the wasm32-wasi target,
which is available as an RPM called rust-std-static-wasm32-wasi.
Closes#12896
[avi: regenerate toolchain]
Closes#13258
Remove redundant "Total: ..." line.
Include the entire `reader_concurrency_semaphore::stats` in the printout. This includes a lot of metrics not exported to monitoring. These metrics are very valuable when debugging timeouts but are otherwise uninteresting. To avoid bloating our monitoring with such niche metrics, we dump them when they are interesting: when timeouts happen. To be really helpful, we do need historic values too, but this shouldn't be a problem: timeouts come in bursts, we usually get at least a handful of diagnostics dumps at a time.
New stats are also added to record the reason why reads are queued on the semaphore.
Printout before:
```
INFO 2023-03-14 12:43:54,496 [shard 0] reader_concurrency_semaphore - Semaphore test_reader_concurrency_semaphore_memory_limit_no_leaks with 4/4 count and 7168/4096 memory resources: kill limit triggered, dumping permit diagnostics:
permits count memory table/description/state
4 4 7K *.*/reader/active/unused
2 0 0B *.*/reader/waiting_for_admission
6 4 7K total
Total: 6 permits with 4 count and 7K memory resources
```
Printout after:
```
INFO 2023-03-16 04:23:41,791 [shard 0] reader_concurrency_semaphore - Semaphore test_reader_concurrency_semaphore_memory_limit_no_leaks with 3/4 count and 7168/4096 memory resources: kill limit triggered, dumping permit diagnostics:
permits count memory table/description/state
2 2 6K *.*/reader/active/unused
1 1 1K *.*/reader/waiting_for_memory
2 0 0B *.*/reader/waiting_for_admission
5 3 7K total
Stats:
permit_based_evictions: 0
time_based_evictions: 0
inactive_reads: 0
total_successful_reads: 0
total_failed_reads: 0
total_reads_shed_due_to_overload: 0
total_reads_killed_due_to_kill_limit: 1
reads_admitted: 4
reads_enqueued_for_admission: 4
reads_enqueued_for_memory: 5
reads_admitted_immediately: 2
reads_queued_because_ready_list: 0
reads_queued_because_used_permits: 0
reads_queued_because_memory_resources: 0
reads_queued_because_count_resources: 4
reads_queued_with_eviction: 0
total_permits: 6
current_permits: 5
used_permits: 0
blocked_permits: 0
disk_reads: 0
sstables_read: 0
```
Closes#13173
* github.com:scylladb/scylladb:
test/boost/reader_concurrency_semaphore_test: remove redundant stats printouts
reader_concurrency_semaphore: do_dump_reader_permit_diagnostics(): print the stats
reader_concurrency_semaphore: add stats to record reason for queueing permits
reader_concurrency_semaphore: can_admit_read(): also return reason for rejection
It's name is too generic despite it's narrow specialization. Also,
there's a version_from_string() method that does the same in a more
convenient way.
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
There are two string->{version|format} converters living on class
sstable. It's better to have both in namespace scope. Surprisingly,
there's only one caller of it.
Also this patch makes both accept std::string_view not to limit the
helpers in converting only sstring&-s. This changes calls for
reverse_map template update with "heterogenuous lookup".
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
This way the version type can be fed as-is into fmt:: code, respectively
the conversion to string is as simple as fmt::to_string(v). So also drop
the explicit existing to_string() helper updating all callers.
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
