before this change, we rely on `using namespace seastar` to use
`seastar::format()` without qualifying the `format()` with its
namespace. this works fine until we changed the parameter type
of format string `seastar::format()` from `const char*` to
`fmt::format_string<...>`. this change practically invited
`seastar::format()` to the club of `std::format()` and `fmt::format()`,
where all members accept a templated parameter as its `fmt`
parameter. and `seastar::format()` is not the best candidate anymore.
despite that argument-dependent lookup (ADT for short) favors the
function which is in the same namespace as its parameter, but
`using namespace` makes `seastar::format()` more competitive,
so both `std::format()` and `seastar::format()` are considered
as the condidates.
that is what is happening scylladb in quite a few caller sites of
`format()`, hence ADT is not able to tell which function the winner
in the name lookup:
```
/__w/scylladb/scylladb/mutation/mutation_fragment_stream_validator.cc:265:12: error: call to 'format' is ambiguous
265 | return format("{} ({}.{} {})", _name_view, s.ks_name(), s.cf_name(), s.id());
| ^~~~~~
/usr/bin/../lib/gcc/x86_64-redhat-linux/14/../../../../include/c++/14/format:4290:5: note: candidate function [with _Args = <const std::basic_string_view<char> &, const seastar::basic_sstring<char, unsigned int, 15> &, const seastar::basic_sstring<char, unsigned int, 15> &, const utils::tagged_uuid<table_id_tag> &>]
4290 | format(format_string<_Args...> __fmt, _Args&&... __args)
| ^
/__w/scylladb/scylladb/seastar/include/seastar/core/print.hh:143:1: note: candidate function [with A = <const std::basic_string_view<char> &, const seastar::basic_sstring<char, unsigned int, 15> &, const seastar::basic_sstring<char, unsigned int, 15> &, const utils::tagged_uuid<table_id_tag> &>]
143 | format(fmt::format_string<A...> fmt, A&&... a) {
| ^
```
in this change, we
change all `format()` to either `fmt::format()` or `seastar::format()`
with following rules:
- if the caller expects an `sstring` or `std::string_view`, change to
`seastar::format()`
- if the caller expects an `std::string`, change to `fmt::format()`.
because, `sstring::operator std::basic_string` would incur a deep
copy.
we will need another change to enable scylladb to compile with the
latest seastar. namely, to pass the format string as a templated
parameter down to helper functions which format their parameters.
to miminize the scope of this change, let's include that change when
bumping up the seastar submodule. as that change will depend on
the seastar change.
Signed-off-by: Kefu Chai <kefu.chai@scylladb.com>
Hides the functionality behind a cluster feature, i.e. postspones
using it until an upgrade is complete etc. This to allow rolling back
even with dirty nodes, at least until a cluster is commited.
Feature can also be disabled by scylla option, just in case. This will
lock it out of whole cluster, but this is probably good, because depending
on off or on, certain schema/raft ops might fail or succeed (due to large
mutations), and this should probably be equivalent across nodes.
Bind variables in CQL have two formats: positional (`?`) where a variable is referred to by its relative position in the statement, and named (`:var`), where the user is expected to supply a name->value mapping.
In 19a6e69001 we identified the case where a named bind variable appears twice in a query, and collapsed it to a single entry in the statement metadata. Without this, a driver using the named variable syntax cannot disambiguate which variable is referred to.
However, it turns out that users can use the positional call form even with the named variable syntax, by using the positional API of the driver. To support this use case, we add a configuration variable to disable the same-variable detection.
Because the detection has to happen when the entire statement is visible, we have to supply the configuration to the parser. We call it the `dialect` and pass it from all callers. The alternative would be to add a pre-prepare call similar to fill_prepare_context that rewrites all expressions in a statement to deduplicate variables.
A unit test is added.
Fixes#15559
This may be useful to users transitioning from Cassandra, so merits a backport.
Closesscylladb/scylladb#19493
* github.com:scylladb/scylladb:
cql3: add option to not unify bind variables with the same name
cql3: introduce dialect infrastructure
cql3: prepared_statement_cache: drop cache key default constructor
Bind variables in CQL have two formats: positional (`?`) where a
variable is referred to by its relative position in the statement,
and named (`:var`), where the user is expected to supply a
name->value mapping.
In 19a6e69001 we identified the case where a named bind variable
appears twice in a query, and collapsed it to a single entry in the
statement metadata. Without this, a driver using the named variable
syntax cannot disambiguate which variable is referred to.
However, it turns out that users can use the positional call form
even with the named variable syntax, by using the positional
API of the driver. To support this use case, we add a configuration
variable to disable the same-variable detection.
Because the detection has to happen when the entire statement is
visible, we have to supply the configuration to the parser. We
call it the `dialect` and pass it from all callers. The alternative
would be to add a pre-prepare call similar to fill_prepare_context that
rewrites all expressions in a statement to deduplicate variables.
A unit test is added.
Fixes#15559
We revive the `join_ring` option. We support it only in the
Raft-based topology, as we plan to remove the gossip-based topology
when we fix the last blocker - the implementation of the manual
recovery tool. In the Raft-based topology, a node can be assigned
tokens only once when it joins the cluster. Hence, we disallow
joining the ring later, which is possible in Cassandra.
The main idea behind the solution is simple. We make the unsupported
special case of zero tokens a supported normal case. Nodes with zero
tokens assigned are called "zero-token nodes" from now on.
From the topology point of view, zero-token nodes are the same as
token-owning nodes. They can be in the same states, etc. From the
data point of view, they are different. They are not members of
the token ring, so they are not present in
`token_metadata::_normal_token_owners`. Hence, they are ignored in
all non-local replication strategies. The tablet load balancer also
ignores them.
Topology operations involving zero-token nodes are simplified:
- `add` and `replace` finish in the `join_group0` state, so creating
a new CDC generation and streaming are skipped,
- `removenode` and `decommission` skip streaming,
- `rebuild` does not even contact the topology coordinator as there
is nothing to rebuild,
Also, if the topology operation involves a token-owning node,
zero-token nodes are ignored in streaming.
Zero-token nodes can be used as coordinator-only nodes, just like in
Cassandra. They can handle requests just like token-owning nodes.
The main motivation behind zero-token nodes is that they can prevent
the Raft majority loss efficiently. Zero-token nodes are group 0
voters, but they can run on much weaker and cheaper machines because
they do not replicate data and handle client requests by default
(drivers ignore them). For example, if there are two DCs, one with 4
nodes and one with 5 nodes, if we add a DC with 2 zero-token nodes,
every DC will contain less than half of the nodes, so we won't lose
the majority when any DC dies.
Another way of preventing the Raft majority loss is changing the
voter set, which is tracked by scylladb/scylladb#18793. That approach
can be used together with zero-token nodes. In the example above, if
we choose equal numbers of voters in both DCs, then a DC with one
zero-token node will be sufficient. However, in the typical setup of
2 DCs with the same number of nodes it is enough to add a DC with
only one zero-token node without changing the voter set.
Zero-token nodes could also be used as load balancers in the
Alternator.
Reverse reads have already been with us for a while, thus this back
door option to read entire paritions forward and reversing them after
can be retired.
When you SELECT a boolean from system.config, it reads as true/false, but this isn't accepted
on UPDATE (instead, we accept 1/0). This is surprising and annoying, so accept true/false in
both directions.
Not a regression, so a backport isn't strictly necessary.
Closesscylladb/scylladb#19792
* github.com:scylladb/scylladb:
config: specialize from-string conversion for bool
config: wrap boost::lexical_cast<> when converting from strings
If set, any remaining segment that has data older than this threshold will request flushing, regardless of data pressure. I.e. even a system where nothing happends will after X seconds flush data to free up the commit log.
Related to #15820
The functionality here is to prevent pathological/test cases where a silent system cannot fully process stuff like compaction, GC etc due to things like CL forcing smaller GC windows etc.
Closesscylladb/scylladb#15971
* github.com:scylladb/scylladb:
commitlog: Make max data lifetime runtime-configurable
db::config: Expose commitlog_max_data_lifetime_in_s parameter
commitlog: Add optional max lifetime parameter to cl instance
The yaml/json representation for bool is true/false, but boost::lexical_cast
is 1/0. Specialize bool conversion to accept true/false (for yaml/json
compatibilty) and 1/0 (for backward compatibility). This provides
round-trip conversion for bool configs in system.config.
Configuration uses boost::lexical_cast to convert strings to native
values (e.g. bools/ints). However, boost::lexical_cast doesn't
recognize true/false for bool. Since we can't change boost::lexical_cast,
replace it with a wrapper that forwards directly to boost::lexical_cast.
In the next step, we'll specialize it for bool.
We currently disable tombstone GC for compaction done on the read path of streaming and repair, because those expired tombstones can still prevent data resurrection. With time-based tombstone GC, missing a repair for long enough can cause data resurrection because a tombstone is potentially GC'd before it could be spread to every node by repair. So repair disseminating these expired tombstones helps clusters which missed repair for long enough. It is not a guarantee because compaction could have done the GC itself, but it is better than nothing.
This last resort is getting less important with repair-based tombstone GC. Furthermore, we have seen this cause huge repair amplification in a cluster, where expired tombstones triggered repair replicating otherwise identical rows.
This series makes tombstone GC on the streaming/repair compaction path configurable with a config item. This new config item defaults to `false` (current behaviour), setting it to `true`, will enable tombstone GC.
Fixes: https://github.com/scylladb/scylladb/issues/19015
Not a regression, no backport needed
Closesscylladb/scylladb#19016
* github.com:scylladb/scylladb:
test/topology_custom/test_repair: add test for enable_tombstone_gc_for_streaming_and_repair
replica/table: maybe_compact_for_streaming(): toggle tombstone GC based on the control flag
replica: propagate enable_tombstone_gc_for_streaming_and_repair to maybe_compact_for_streaming()
db/config: introduce enable_tombstone_gc_for_streaming_and_repair
The reader concurrency semaphore restricts the concurrency of reads that require CPU (intention: they read from the cache) to 1, meaning that if there is even a single active read which declares that it needs just CPU to proceed, no new read is admitted. This is meant to keep the concurrency of reads in the cache at 1. The idea is that concurrency in the cache is not useful: it just leads to the reactor rotating between these reads, all of the finishing later then they could if they were the only active read in the cache.
This was observed to backfire in the case where there reads from a single table are mostly very fast, but on some keys are very slow (hint: collection full of tombstones). In this case the slow read keeps up the fast reads in the queue, increasing the 99th percentile latencies significantly.
This series proposes to fix this, by making the CPU concurrency configurable. We don't like tunables like this and this is not a proper fix, but a workaround. The proper fix would be to allow to cut any page early, but we cannot cut a page in the middle of a row. We could maybe have a way of detecting slow reads and excluding them from the CPU concurrency. This would be a heuristic and it would be hard to get right. So in this series a robust and simple configurable is offered, which can be used on those few clusters which do suffer from the too strict concurrency limit. We have seen it in very few cases so far, so this doesn't seem to be wide-spread.
Fixes: https://github.com/scylladb/scylladb/issues/19017
This fixes a regression introduced in 5.0, so we have to backport to all currently supported releases
Closesscylladb/scylladb#19018
* github.com:scylladb/scylladb:
test/boost/reader_concurrency_semaphore_test: add test for live-configurable cpu concurrenc Please enter the commit message for your changes. Lines starting
test/boost/reader_concurrency_semaphore_test: hoist require_can_admit
reader_concurrency_semaphore: wire in the configurable cpu concurrency
reader_concurrency_semaphore: add cpu_concurrency constructor parameter
db/config: introduce reader_concurrency_semahore_cpu_concurrency
To control whether the compacting reader (if enabled) for streaming and
repair can garbage-collect tombstones.
Default is false (previous behaviour).
Not wired yet.
seastar::format() does not use operator<< under the hood, it uses
{fmt}, so update the comment accordingly.
Signed-off-by: Kefu Chai <kefu.chai@scylladb.com>
Closesscylladb/scylladb#19315
Making the count resources on the maintenance (streaming) semaphore live update via config. This will allow us to improve repair speed on mixed-shard clusters, where we suspect that reader trashing -- due to the combination of high number of readers on each shard and very conservative reader count limit (10) -- is the main cause of the slowness.
Making this count limit confgurable allows us to start experimenting with this fix, without committing to a count limit increase (or removal), addressing the pain in the field.
Refs: #18269
No OSS backport needed.
Closesscylladb/scylladb#19248
* github.com:scylladb/scylladb:
replica/database: wire in maintenance_reader_concurrency_semaphore_count_limit
db/config: introduce maintenance_reader_concurrency_semaphore_count_limit
reader_concurrency_semaphore: make count parameter live-update
thrift support was deprecated since ScyllaDB 5.2
> Thrift API - legacy ScyllaDB (and Apache Cassandra) API is
> deprecated and will be removed in followup release. Thrift has
> been disabled by default.
so let's drop it. in this change,
* thrift protocol support is dropped
* all references to thrift support in document are dropped
* the "thrift_version" column in system.local table is preserved for backward compatibility, as we could load from an existing system.local table which still contains this clolumn, so we need to write this column as well.
* "/storage_service/rpc_server" is only preserved for backward compatibility with java-based nodetool.
Fixes#3811Fixes#18416
Signed-off-by: Kefu Chai <kefu.chai@scylladb.com>
- [x] not a fix, no need to backport
Closesscylladb/scylladb#18453
* github.com:scylladb/scylladb:
config: expand on rpc_keepalive's description
api: s/rpc/thrift/
db/system_keyspace: drop thrift_version from system.local table
transport: do not return client_type from cql_server::connection::make_client_key()
treewide: drop thrift support
This features used to be there for a while, but then it was removed by
83d491af02. This patch partially takes it
back, but maps to UNUSED, so that if met in config, it's warned, but
other features are parsed as well.
refs: #18968
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
before this change, we use "RPC or native". but before thrift support
is removed "RPC" implies "thrift", now that we've dropped thrift
support, "RPC" could be confusing here, so let's be more specific,
and put all connection types in place of "RPC or native".
Signed-off-by: Kefu Chai <kefu.chai@scylladb.com>
thrift support was deprecated since ScyllaDB 5.2
> Thrift API - legacy ScyllaDB (and Apache Cassandra) API is
> deprecated and will be removed in followup release. Thrift has
> been disabled by default.
so let's drop it. in this change,
* thrift protocol support is dropped
* all references to thrift support in document are dropped
* the "thrift_version" column in system.local table is
preserved for backward compatibility, as we could load
from an existing system.local table which still contains
this clolumn, so we need to write this column as well.
* "/storage_service/rpc_server" is only preserved for
backward compatibility with java-based nodetool.
* `rpc_port` and `start_rpc` options are preserved, but
they are marked as "Unused". so that the new release
of scylladb can consume existing scylla.yaml configurations
which might contain these settings. by making them
deprecated, user will be able get warned, and update
their configurations before we actually remove them
in the next major release.
Fixes#3811Fixes#18416
Signed-off-by: Kefu Chai <kefu.chai@scylladb.com>
Incremented the components_memory_reclaim_threshold config's default
value to 0.2 as the previous value was too strict and caused unnecessary
eviction in otherwise healthy clusters.
Fixes#18607
Signed-off-by: Lakshmi Narayanan Sreethar <lakshmi.sreethar@scylladb.com>
Closesscylladb/scylladb#18964
... and replace it with boolean enable_tablets option. All the places
in the code are patched to check the latter option instead of the former
feature.
The option is OFF by default, but the default scylla.yaml file sets this
to true, so that newly installed clusters turn tablets ON.
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
Closesscylladb/scylladb#18898
since we've migrated away from the generic homebrew formatters
for range-alike containers, there is no need to keep there operator<<
around -- they were preserved in order to work with the container
formatters which expect operator<< of the elements.
Signed-off-by: Kefu Chai <kefu.chai@scylladb.com>
In commit 642f9a1966 (repair: Improve
estimated_partitions to reduce memory usage), a 10% hard coded
estimation ratio is used.
This patch introduces a new config option to specify the estimation
ratio of partitions written by repair out of the total partitions.
It is set to 0.1 by default.
Fixes#18615Closesscylladb/scylladb#18634
The direct failure detector design is simplistic. It sends pings
sequentially and times out listeners that reached the threshold (i.e.
didn't hear from a given endpoint for too long) in-between pings.
Given the sequential nature, the previous ping must finish so the next
ping can start. We timeout pings that take too long. The timeout was
hardcoded and set to 300ms. This is too low for wide-area setups --
latencies across the Earth can indeed go up to 300ms. 3 subsequent timed
out pings to a given node were sufficient for the Raft listener to "mark
server as down" (the listener used a threshold of 1s).
Increase the ping timeout to 600ms which should be enough even for
pinging the opposite side of Earth, and make it tunable.
Increase the Raft listener threshold from 1s to 2s. Without the
increased threshold, one timed out ping would be enough to mark the
server as down. Increasing it to 2s requires 3 timed out pings which
makes it more robust in presence of transient network hiccups.
In the future we'll most likely want to decrease the Raft listener
threshold again, if we use Raft for data path -- so leader elections
start quickly after leader failures. (Faster than 2s). To do that we'll
have to improve the design of the direct failure detector.
Ref: scylladb/scylladb#16410Fixes: scylladb/scylladb#16607
---
I tested the change manually using `tc qdisc ... netem delay`, setting
network delay on local setup to ~300ms with jitter. Without the change,
the result is as observed in scylladb/scylladb#16410: interleaving
```
raft_group_registry - marking Raft server ... as dead for Raft groups
raft_group_registry - marking Raft server ... as alive for Raft groups
```
happening once every few seconds. The "marking as dead" happens whenever
we get 3 subsequent failed pings, which is happens with certain (high)
probability depending on the latency jitter. Then as soon as we get a
successful ping, we mark server back as alive.
With the change, the phenomenon no longer appears.
Closesscylladb/scylladb#18443
We move consistent cluster management out of experimental and
make it the default for new clusters in 6.0. In code, we make the
`consistent-topology-changes` flag unused and assumed to be true.
In 6.0, the topology upgrade procedure will be manual and
voluntary, so some clusters will still be using the gossip-based
topology even though they support the raft-based topology.
Therefore, we need to continue testing the gossip-based topology.
This is possible by using the `force-gossip-topology-changes` flag
introduced in scylladb/scylladb#18284.
Ref scylladb/scylladb#17802Closesscylladb/scylladb#18285
* github.com:scylladb/scylladb:
docs: raft.rst: update after removing consistent-topology-changes
treewide: fix indentation after the previous patch
db: config: make consistent-topology-changes unused
test: lib: single_node_cql_env: restart a node in noninitial run_in_thread calls
test: test_read_required_hosts: run with force-gossip-topology-changes
storage_service: join_cluster: replace force_gossip_based_join with force-gossip-topology-changes
storage_service: join_token_ring: fix finish_setup_after_join calls
Repair may miss some tablets that migrated across nodes.
So if tombstones expire after some timeout, then we can
have data resurrection.
Set default tombstone_gc mode to "repair" for tables which
use tablets (if repair is required).
Fixes: #16627.
Closesscylladb/scylladb#18013
* github.com:scylladb/scylladb:
test: check default value of tombstone_gc
test: topology: move some functions to util.py
cql3: statements: change default tombstone_gc mode for tablets
Metric family config lets a user configure the metric family aggregate labels.
This patch modifies the existing relable-config from file to accept
metric family config.
Similar to the existing relable_config, it adds a metric_family_configs
section. For example, the following configuration demonstrates changing
aggregate labels by name and regular expression.
```
metric_family_configs:
- name: storage_service
aggregate_labels: [shard]
- regex: (storage_proxy.*)
aggregate_labels: [shard, scheduling_group_name]
```
Signed-off-by: Amnon Heiman <amnon@scylladb.com>
Closesscylladb/scylladb#18339
We make the `consistent-topology-changes` experimental feature
unused and assumed to be true in 6.0. We remove code branches that
executed if `consistent-topology-changes` was disabled.
Currently, if tombstone_gc mode isn't specified for a table,
then "timeout" is used by default. With tablets, running
"nodetool repair -pr" may miss a tablet if it migrated across
the nodes. Then, if we expire tombstones for ranges that
weren't repaired, we may get data resurrection.
Set default tombstone_gc mode value for DDLs that don't
specify it. It's set to "repair" for tables which use tablets
unless they use local replication strategy or rf = 1.
Otherwise it's set to "timeout".
We are going to make the `consistent-topology-changes` experimental
feature unused in 6.0. However, the topology upgrade procedure will
be manual and voluntary, so some 6.0 clusters will be using the
gossip-based topology. Therefore, we need to continue testing the
gossip-based topology. The solution is introducing a new flag,
`force-gossip-topology-changes`, that will enforce the gossip-based
topology in a fresh cluster.
In this patch, we only introduce the parameter without any effect.
Here is the explanation. Making `consistent-topology-changes` unused
and introducing `force-gossip-topology-changes` requires adjustments
in scylla-dtest. We want to merge changes to scylladb and scylla-dtest
in a way that ensures all tests are run correctly during the whole
process. If we merged all changes to scylladb first, before merging
the scylla-dtest changes, all tests would run with the raft-based
topology and the ones excluded in the raft-based topology would fail.
We also can't merge all changes to scylla-dtest first. However, we
can follow this plan:
1. scylladb: merge this patch
2. scylla-dtest: start using `force-gossip-topology-changes`
in jobs that run without the raft-based topology
3. scylladb: merge the rest of the changes
4. scylla-dtest: merge the rest of the changes
Ref scylladb/scylladb#17802Closesscylladb/scylladb#18284
before this change, we rely on the default-generated fmt::formatter
created from operator<<, but fmt v10 dropped the default-generated
formatter.
in this change, we include `fmt/ranges.h` and/or `fmt/std.h`
for formatting the container types, like vector, map
optional and variant using {fmt} instead of the homebrew
formatter based on operator<<.
with this change, the changes adding fmt::formatter and
the changes using ostream formatter explicitly, we are
allowed to drop `FMT_DEPRECATED_OSTREAM` macro.
Refs scylladb#13245
Signed-off-by: Kefu Chai <kefu.chai@scylladb.com>
A new configuration variable, components_memory_reclaim_threshold, has
been added to configure the maximum allowed percentage of available
memory for all SSTable components in a shard. If the total memory usage
exceeds this threshold, it will be reclaimed from the components to
bring it back under the limit. Currently, only the memory used by the
bloom filters will be restricted.
Signed-off-by: Lakshmi Narayanan Sreethar <lakshmi.sreethar@scylladb.com>
Injection parameters can be used in the lambda passed to
inject_with_handler method to take some values from
the test. However, there was no way to set values to these
parameters on node startup, only through
the error injection REST api. Therefore, we couldn't rely
on this when inject_with_handler is used during
node startup, it could trigger before we call the api
from the test.
In this commit with solve this problem by allowing these
parameters to be assigned through scylla.yaml config.
The defer.hh header was added to error_injection.hh to fix
compilation after adding error_injection.hh to config.hh,
defer function is used in error_injection.hh.
Regulates the page size in bytes via config, instead of the currently
used hard-coded constant. Allows tests to configure lower limits so they
can work with smaller data-sets when testing paging related
functionality.
Not wired yet.
When topology barrier is blocked for longer than configured threshold
(2s), stale versions are marked as stalled and when they get released
they report backtrace to the logs. This should help to identify what
was holding for token metadata pointer for too long.
Example log:
token_metadata - topology version 30 held for 299.159 [s] past expiry, released at: 0x2397ae1 0x23a36b6 ...
Closesscylladb/scylladb#17427
Set filesystem permissions for the maintenance socket to 660 (previously it was 755) to allow a scyllaadm's group to connect.
Split the logic of creating sockets into two separate functions, one for each case: when it is a regular cql controller or used by maintenance_socket.
Fixes https://github.com/scylladb/scylladb/issues/16487.
Closesscylladb/scylladb#17113
* github.com:scylladb/scylladb:
maintenance_socket: add option to set owning group
transport/controller: get rid of magic number for socket path's maximal length
transport/controller: set unix_domain_socket_permissions for maintenance_socket
transport/controller: pass unix_domain_socket_permissions to generic_server::listen
transport/controller: split configuring sockets into separate functions
Option `maintenance-socket-group` sets the owning group of the maintenance socket.
If not set, the group will be the same as the user running the scylla node.
before this change, we rely on the default-generated fmt::formatter
created from operator<<, but fmt v10 dropped the default-generated
formatter.
in this change, we define formatters for `error_injection_at_startup`,
and drop its operator<<.
Refs #13245
Signed-off-by: Kefu Chai <kefu.chai@scylladb.com>
Closesscylladb/scylladb#17211
This reverts commit 370fbd346c, reversing
changes made to 0912d2a2c6.
This makes scylla-manager mis-interpret the data_file_directories
somehow, issue #17078
The motivation for tablet resizing is that we want to keep the average tablet size reasonable, such that load rebalancing can remain efficient. Too large tablet makes migration inefficient, therefore slowing down the balancer.
If the avg size grows beyond the upper bound (split threshold), then balancer decides to split. Split spans all tablets of a table, due to power-of-two constraint.
Likewise, if the avg size decreases below the lower bound (merge threshold), then merge takes place in order to grow the avg size. Merge is not implemented yet, although this series lays foundation for it to be impĺemented later on.
A resize decision can be revoked if the avg size changes and the decision is no longer needed. For example, let's say table is being split and avg size drops below the target size (which is 50% of split threshold and 100% of merge one). That means after split, the avg size would drop below the merge threshold, causing a merge after split, which is wasteful, so it's better to just cancel the split.
Tablet metadata gains 2 new fields for managing this:
resize_type: resize decision type, can be either of "merge", "split", or "none".
resize_seq_number: a sequence number that works as the global identifier of the decision (monotonically increasing, increased by 1 on every new decision emitted by the coordinator).
A new RPC was implemented to pull stats from each table replica, such that load balancer can calculate the avg tablet size and know the "split status", for a given table. Avg size is aggregated carefully while taking RF of each DC into account (which might differ).
When a table is done splitting its storage, it loads (mirror) the resize_seq_number from tablet metadata into its local state (in another words, my split status is ready). If a table is split ready, coordinator will see that table's seq number is the same as the one in tablet metadata. Helps to distinguish stale decisions from the latest one (in case decisions are revoked and re-emited later on). Also, it's aggregated carefully, by taking the minimum among all replicas, so coordinator will only update topology when all replicas are ready.
When load balancer emits split decision, replicas will listen to need to split with a "split monitor" that is awakened once a table has replication metadata updated and detects the need for split (i.e. resize_type field is "split").
The split monitor will start splitting of compaction groups (using mechanism introduced here: 081f30d149) for the table. And once splitting work is completed, the table updates its local state as having completed split.
When coordinator pulls the split status of all replicas for a table via RPC, the balancer can see whether that table is ready for "finalizing" the decision, which is about updating tablet metadata to split each tablet into two. Once table replicas have their replication metadata updated with the new tablet count, they can update appropriately their set of compaction groups (that were previously split in the preparation step).
Fixes#16536.
Closesscylladb/scylladb#16580
* github.com:scylladb/scylladb:
test/topology_experimental_raft: Add tablet split test
replica: Bypass reshape on boot with tablets temporarily
replica: Fix table::compaction_group_for_sstable() for tablet streaming
test/topology_experimental_raft: Disable load balancer in test fencing
replica: Remap compaction groups when tablet split is finalized
service: Split tablet map when split request is finalized
replica: Update table split status if completed split compaction work
storage_service: Implement split monitor
topology_cordinator: Generate updates for resize decisions made by balancer
load_balancer: Introduce metrics for resize decisions
db: Make target tablet size a live-updateable config option
load_balancer: Implement resize decisions
service: Wire table_resize_plan into migration_plan
service: Introduce table_resize_plan
tablet_mutation_builder: Add set_resize_decision()
topology_coordinator: Wire load stats into load balancer
storage_service: Allow tablet split and migration to happen concurrently
topology_coordinator: Periodically retrieve table_load_stats
locator: Introduce topology::get_datacenter_nodes()
storage_service: Implement table_load_stats RPC
replica: Expose table_load_stats in table
replica: Introduce storage_group::live_disk_space_used()
locator: Introduce table_load_stats
tablets: Add resize decision metadata to tablet metadata
locator: Introduce resize_decision
This change is intended to ensure, that
db::config fields related to directories
are not changed. To achieve that a member
function called setup_directories() is
removed.
The responsibility for directories paths
has been moved to utils::directories,
which may generate default paths if the
configuration does not provide a specific
value.
Fixes: scylladb#5626
Signed-off-by: Patryk Wrobel <patryk.wrobel@scylladb.com>
