Files
scylladb/replica/tablets.hh
Benny Halevy 3feb759943 everywhere: use utils::chunked_vector for list of mutations
Currently, we use std::vector<*mutation> to keep
a list of mutations for processing.
This can lead to large allocation, e.g. when the vector
size is a function of the number of tables.

Use a chunked vector instead to prevent oversized allocations.

`perf-simple-query --smp 1` results obtained for fixed 400MHz frequency
and PGO disabled:

Before (read path):
```
enable-cache=1
Running test with config: {partitions=10000, concurrency=100, mode=read, query_single_key=no, counters=no}
Disabling auto compaction
Creating 10000 partitions...

89055.97 tps ( 66.1 allocs/op,   0.0 logallocs/op,  14.2 tasks/op,   39417 insns/op,   18003 cycles/op,        0 errors)
103372.72 tps ( 66.1 allocs/op,   0.0 logallocs/op,  14.2 tasks/op,   39380 insns/op,   17300 cycles/op,        0 errors)
98942.27 tps ( 66.1 allocs/op,   0.0 logallocs/op,  14.2 tasks/op,   39413 insns/op,   17336 cycles/op,        0 errors)
103752.93 tps ( 66.1 allocs/op,   0.0 logallocs/op,  14.2 tasks/op,   39407 insns/op,   17252 cycles/op,        0 errors)
102516.77 tps ( 66.1 allocs/op,   0.0 logallocs/op,  14.2 tasks/op,   39403 insns/op,   17288 cycles/op,        0 errors)
throughput:
	mean=   99528.13 standard-deviation=6155.71
	median= 102516.77 median-absolute-deviation=3844.59
	maximum=103752.93 minimum=89055.97
instructions_per_op:
	mean=   39403.99 standard-deviation=14.25
	median= 39406.75 median-absolute-deviation=9.30
	maximum=39416.63 minimum=39380.39
cpu_cycles_per_op:
	mean=   17435.81 standard-deviation=318.24
	median= 17300.40 median-absolute-deviation=147.59
	maximum=18002.53 minimum=17251.75
```

After (read path)
```
enable-cache=1
Running test with config: {partitions=10000, concurrency=100, mode=read, query_single_key=no, counters=no}
Disabling auto compaction
Creating 10000 partitions...
59755.04 tps ( 66.2 allocs/op,   0.0 logallocs/op,  14.2 tasks/op,   39466 insns/op,   22834 cycles/op,        0 errors)
71854.16 tps ( 66.1 allocs/op,   0.0 logallocs/op,  14.2 tasks/op,   39417 insns/op,   17883 cycles/op,        0 errors)
82149.45 tps ( 66.1 allocs/op,   0.0 logallocs/op,  14.2 tasks/op,   39411 insns/op,   17409 cycles/op,        0 errors)
49640.04 tps ( 66.1 allocs/op,   0.0 logallocs/op,  14.3 tasks/op,   39474 insns/op,   19975 cycles/op,        0 errors)
54963.22 tps ( 66.1 allocs/op,   0.0 logallocs/op,  14.3 tasks/op,   39474 insns/op,   18235 cycles/op,        0 errors)
throughput:
	mean=   63672.38 standard-deviation=13195.12
	median= 59755.04 median-absolute-deviation=8709.16
	maximum=82149.45 minimum=49640.04
instructions_per_op:
	mean=   39448.38 standard-deviation=31.60
	median= 39466.17 median-absolute-deviation=25.75
	maximum=39474.12 minimum=39411.42
cpu_cycles_per_op:
	mean=   19267.01 standard-deviation=2217.03
	median= 18234.80 median-absolute-deviation=1384.25
	maximum=22834.26 minimum=17408.67
```

`perf-simple-query --smp 1 --write` results obtained for fixed 400MHz frequency
and PGO disabled:

Before (write path):
```
enable-cache=1
Running test with config: {partitions=10000, concurrency=100, mode=write, query_single_key=no, counters=no}
Disabling auto compaction
63736.96 tps ( 59.4 allocs/op,  16.4 logallocs/op,  14.3 tasks/op,   49667 insns/op,   19924 cycles/op,        0 errors)
64109.41 tps ( 59.3 allocs/op,  16.0 logallocs/op,  14.3 tasks/op,   49992 insns/op,   20084 cycles/op,        0 errors)
56950.47 tps ( 59.3 allocs/op,  16.0 logallocs/op,  14.3 tasks/op,   50005 insns/op,   20501 cycles/op,        0 errors)
44858.42 tps ( 59.3 allocs/op,  16.0 logallocs/op,  14.3 tasks/op,   50014 insns/op,   21947 cycles/op,        0 errors)
28592.87 tps ( 59.3 allocs/op,  16.0 logallocs/op,  14.3 tasks/op,   50027 insns/op,   27659 cycles/op,        0 errors)
throughput:
	mean=   51649.63 standard-deviation=15059.74
	median= 56950.47 median-absolute-deviation=12087.33
	maximum=64109.41 minimum=28592.87
instructions_per_op:
	mean=   49941.18 standard-deviation=153.76
	median= 50005.24 median-absolute-deviation=73.01
	maximum=50027.07 minimum=49667.05
cpu_cycles_per_op:
	mean=   22023.01 standard-deviation=3249.92
	median= 20500.74 median-absolute-deviation=1938.76
	maximum=27658.75 minimum=19924.32
```

After (write path)
```
enable-cache=1
Running test with config: {partitions=10000, concurrency=100, mode=write, query_single_key=no, counters=no}
Disabling auto compaction
53395.93 tps ( 59.4 allocs/op,  16.5 logallocs/op,  14.3 tasks/op,   50326 insns/op,   21252 cycles/op,        0 errors)
46527.83 tps ( 59.3 allocs/op,  16.0 logallocs/op,  14.3 tasks/op,   50704 insns/op,   21555 cycles/op,        0 errors)
55846.30 tps ( 59.3 allocs/op,  16.0 logallocs/op,  14.3 tasks/op,   50731 insns/op,   21060 cycles/op,        0 errors)
55669.30 tps ( 59.3 allocs/op,  16.0 logallocs/op,  14.3 tasks/op,   50735 insns/op,   21521 cycles/op,        0 errors)
52130.17 tps ( 59.3 allocs/op,  16.0 logallocs/op,  14.3 tasks/op,   50757 insns/op,   21334 cycles/op,        0 errors)
throughput:
	mean=   52713.91 standard-deviation=3795.38
	median= 53395.93 median-absolute-deviation=2955.40
	maximum=55846.30 minimum=46527.83
instructions_per_op:
	mean=   50650.57 standard-deviation=182.46
	median= 50731.38 median-absolute-deviation=84.09
	maximum=50756.62 minimum=50325.87
cpu_cycles_per_op:
	mean=   21344.42 standard-deviation=202.86
	median= 21334.00 median-absolute-deviation=176.37
	maximum=21554.61 minimum=21060.24
```

Fixes #24815

Improvement for rare corner cases. No backport required

Signed-off-by: Benny Halevy <bhalevy@scylladb.com>

Closes scylladb/scylladb#24919
2025-07-13 19:13:11 +03:00

112 lines
4.2 KiB
C++

/*
* Copyright (C) 2023-present ScyllaDB
*/
/*
* SPDX-License-Identifier: LicenseRef-ScyllaDB-Source-Available-1.0
*/
#pragma once
#include "types/types.hh"
#include "timestamp.hh"
#include "locator/tablets.hh"
#include "schema/schema_fwd.hh"
#include "mutation/mutation.hh"
#include "mutation/canonical_mutation.hh"
#include "replica/database_fwd.hh"
#include <seastar/core/future.hh>
#include <vector>
namespace cql3 {
class query_processor;
}
namespace gms {
class feature_service;
}
namespace replica {
data_type get_replica_set_type();
data_type get_tablet_info_type();
schema_ptr make_tablets_schema();
void tablet_add_repair_scheduler_user_types(const sstring& ks, replica::database& db);
std::vector<data_value> replicas_to_data_value(const locator::tablet_replica_set& replicas);
/// Converts information in tablet_map to mutations of system.tablets.
///
/// The mutations will delete any older tablet information for the same table.
/// The provided timestamp should be strictly monotonically increasing
/// between calls for the overriding to work correctly.
future<mutation> tablet_map_to_mutation(const locator::tablet_map&,
table_id,
const sstring& keyspace_name,
const sstring& table_name,
api::timestamp_type,
const gms::feature_service& features);
mutation colocated_tablet_map_to_mutation(table_id,
const sstring& keyspace_name,
const sstring& table_name,
table_id base_table,
api::timestamp_type);
mutation make_drop_tablet_map_mutation(table_id, api::timestamp_type);
/// Stores a given tablet_metadata in system.tablets.
///
/// Overrides tablet maps for tables present in the given tablet metadata.
/// Does not delete tablet maps for tables which are absent in the given tablet metadata.
/// The provided timestamp should be strictly monotonically increasing
/// between calls for tablet map overriding to work correctly.
/// The timestamp must be greater than api::min_timestamp.
future<> save_tablet_metadata(replica::database&, const locator::tablet_metadata&, api::timestamp_type);
/// Extract a tablet metadata change hint from the tablet mutations.
///
/// Mutations which don't mutate the tablet table are ignored.
std::optional<locator::tablet_metadata_change_hint> get_tablet_metadata_change_hint(const utils::chunked_vector<canonical_mutation>&);
/// Update the tablet metadata change hint, with the changes represented by the tablet mutation.
///
/// If the mutation belongs to another table, no updates are done.
void update_tablet_metadata_change_hint(locator::tablet_metadata_change_hint&, const mutation&);
/// Reads the replica set from given cell value
locator::tablet_replica_set tablet_replica_set_from_cell(const data_value&);
/// Reads tablet metadata from system.tablets.
future<locator::tablet_metadata> read_tablet_metadata(cql3::query_processor&);
/// Reads the set of hosts referenced by tablet replicas.
future<std::unordered_set<locator::host_id>> read_required_hosts(cql3::query_processor&);
/// Update tablet metadata from system.tablets, based on the provided hint.
///
/// The hint is used to determine what has changed and only reload the changed
/// parts from disk, updating the passed-in metadata in-place accordingly.
future<> update_tablet_metadata(replica::database& db, cql3::query_processor&, locator::tablet_metadata&, const locator::tablet_metadata_change_hint&);
/// Reads tablet metadata from system.tablets in the form of mutations.
future<utils::chunked_vector<canonical_mutation>> read_tablet_mutations(seastar::sharded<database>&);
/// Reads tablet transition stage (if any)
future<std::optional<locator::tablet_transition_stage>> read_tablet_transition_stage(cql3::query_processor& qp, table_id tid, dht::token last_token);
/// Validates changes to system.tablets represented by mutations
void validate_tablet_metadata_change(const locator::tablet_metadata& tm, const utils::chunked_vector<canonical_mutation>& mutations);
} // namespace replica