mirror of
https://github.com/scylladb/scylladb.git
synced 2026-04-29 12:47:02 +00:00
801bf42ea2
Currently, we use storage_proxy/get_cas_shard -> sharder.shard_for_reads to decide which shard to use for LWT code execution on both replicas and the coordinator. If the coordinator is not a replica, shard_for_reads returns 0 — the 'default' shard. This behavior has at least two problems: * Shard 0 may become overloaded, because all LWT coordinators that are not replicas will be served on it. * The zero shard does not match shard_for_reads on replicas, which hinders the "same shard for client and server" RPC-level optimization. To fix this, we need to know whether the current node hosts a replica for the tablet corresponding to the given token. Currently, there is no API we could use for this. For historical reasons, sharder::shard_for_reads returns 0 when the node does not host the shard, which leads to ambiguity. This commit introduces try_get_shard_for_reads, which returns a disengaged std::optional when the tablet is not present on the local node. We leave shard_for_reads method in the base sharder class, it calls try_get_shard_for_reads and returns zero by default. We need to rename tablet_sharder private methods shard_for_reads and shard_for_writes so that they don't conflict with the sharder::shard_for_reads.
54 lines
1.9 KiB
C++
54 lines
1.9 KiB
C++
/*
|
|
* Copyright (C) 2023-present ScyllaDB
|
|
*/
|
|
|
|
/*
|
|
* SPDX-License-Identifier: LicenseRef-ScyllaDB-Source-Available-1.0
|
|
*/
|
|
|
|
#pragma once
|
|
|
|
#include "dht/token-sharding.hh"
|
|
#include "locator/abstract_replication_strategy.hh"
|
|
|
|
#include <seastar/core/abort_source.hh>
|
|
#include <seastar/core/shared_ptr.hh>
|
|
|
|
namespace replica {
|
|
class table;
|
|
}
|
|
|
|
namespace dht {
|
|
|
|
/// A sharder which uses table's most up to date effective_replication_map in each invocation.
|
|
/// As a result, the time during which a single token metadata version is kept is reduced.
|
|
/// The purpose of this class is having a sharder which can be passed to long-running processes
|
|
/// which do not need a single view on replication for the whole duration of the process,
|
|
/// and do not need to block topology change barriers.
|
|
///
|
|
/// No guarantees are made about how long token metadata version is kept alive.
|
|
/// It may be released before the next call.
|
|
class auto_refreshing_sharder : public dht::sharder {
|
|
lw_shared_ptr<replica::table> _table;
|
|
locator::effective_replication_map_ptr _erm;
|
|
const dht::sharder* _sharder;
|
|
optimized_optional<seastar::abort_source::subscription> _callback;
|
|
std::optional<write_replica_set_selector> _sel;
|
|
private:
|
|
void refresh();
|
|
public:
|
|
auto_refreshing_sharder(lw_shared_ptr<replica::table> table, std::optional<write_replica_set_selector> sel = std::nullopt);
|
|
|
|
virtual ~auto_refreshing_sharder();
|
|
|
|
virtual std::optional<unsigned> try_get_shard_for_reads(const token& t) const override;
|
|
|
|
virtual dht::shard_replica_set shard_for_writes(const token& t, std::optional<write_replica_set_selector> sel) const override;
|
|
|
|
virtual std::optional<dht::shard_and_token> next_shard_for_reads(const dht::token& t) const override;
|
|
|
|
virtual dht::token token_for_next_shard_for_reads(const dht::token& t, shard_id shard, unsigned spans = 1) const override;
|
|
};
|
|
|
|
} // namespace dht
|