Files
scylladb/db/hints/resource_manager.hh
Piotr Sarna ac7531d8d9 db,hints: decouple in-flight hints limits from resource manager
The resource manager is used to manage common resources between
various hints managers. In-flight hints used to be one of the shared
resources, but it proves to cause starvation, when one manager eats
the whole limit - which may be especially painful if the background
materialized views hints manager starves the regular hints manager,
which can in turn start failing user writes because of admission control.
This patch makes the limit per-manager again,
which effectively reverts the limit to its original behavior.

Fixes #4483
Message-Id: <8498768e8bccbfa238e6a021f51ec0fa0bf3f7f9.1559649491.git.sarna@scylladb.com>
2019-07-12 19:21:26 +03:00

153 lines
5.2 KiB
C++

/*
* Copyright (C) 2018 ScyllaDB
*/
/*
* This file is part of Scylla.
*
* Scylla is free software: you can redistribute it and/or modify
* it under the terms of the GNU Affero General Public License as published by
* the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* Scylla is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with Scylla. If not, see <http://www.gnu.org/licenses/>.
*/
#pragma once
#include <cstdint>
#include <seastar/core/abort_source.hh>
#include <seastar/core/semaphore.hh>
#include <seastar/core/gate.hh>
#include <seastar/core/memory.hh>
#include <seastar/core/future.hh>
#include "seastarx.hh"
#include <unordered_set>
#include <gms/gossiper.hh>
#include "utils/small_vector.hh"
#include "lister.hh"
namespace service {
class storage_proxy;
class storage_service;
}
namespace db {
namespace hints {
future<dev_t> get_device_id(const fs::path& path);
future<bool> is_mountpoint(const fs::path& path);
using timer_clock_type = seastar::lowres_clock;
class manager;
class space_watchdog {
private:
using ep_key_type = gms::inet_address;
static const std::chrono::seconds _watchdog_period;
struct manager_hash {
size_t operator()(const manager& manager) const {
return reinterpret_cast<uintptr_t>(&manager);
}
};
struct manager_comp {
bool operator()(const std::reference_wrapper<manager>& m1, const std::reference_wrapper<manager>& m2) const {
return std::addressof(m1.get()) == std::addressof(m2.get());
}
};
public:
struct per_device_limits {
utils::small_vector<std::reference_wrapper<manager>, 2> managers;
size_t max_shard_disk_space_size;
};
using shard_managers_set = std::unordered_set<std::reference_wrapper<manager>, manager_hash, manager_comp>;
using per_device_limits_map = std::unordered_map<dev_t, per_device_limits>;
private:
size_t _total_size = 0;
shard_managers_set& _shard_managers;
per_device_limits_map& _per_device_limits_map;
future<> _started = make_ready_future<>();
seastar::abort_source _as;
int _files_count = 0;
public:
space_watchdog(shard_managers_set& managers, per_device_limits_map& per_device_limits_map);
void start();
future<> stop() noexcept;
private:
/// \brief Check that hints don't occupy too much disk space.
///
/// Verifies that all \ref manager::_hints_dir dirs for all managers occupy less than \ref resource_manager::max_shard_disk_space_size.
///
/// If they do, stop all end point managers that have more than one hints file - we don't want some DOWN Node to
/// prevent hints to other Nodes from being generated (e.g. due to some temporary overload and timeout).
///
/// This is a simplistic implementation of a manager for a limited shared resource with a minimum guaranteed share for all
/// participants.
///
/// This implementation guarantees at least a single hint share for all end point managers.
void on_timer();
/// \brief Scan files in a single end point directory.
///
/// Add sizes of files in the directory to _total_size. If number of files is greater than 1 add this end point ID
/// to _eps_with_pending_hints so that we may block it if _total_size value becomes greater than the maximum allowed
/// value.
///
/// \param path directory to scan
/// \param ep_name end point ID (as a string)
/// \return future that resolves when scanning is complete
future<> scan_one_ep_dir(fs::path path, manager& shard_manager, ep_key_type ep_key);
};
class resource_manager {
const size_t _max_send_in_flight_memory;
const size_t _min_send_hint_budget;
seastar::semaphore _send_limiter;
space_watchdog::shard_managers_set _shard_managers;
space_watchdog::per_device_limits_map _per_device_limits_map;
space_watchdog _space_watchdog;
public:
static constexpr size_t hint_segment_size_in_mb = 32;
static constexpr size_t max_hints_per_ep_size_mb = 128; // 4 files 32MB each
static constexpr size_t max_hints_send_queue_length = 128;
public:
resource_manager(size_t max_send_in_flight_memory)
: _max_send_in_flight_memory(std::max(max_send_in_flight_memory, max_hints_send_queue_length))
, _min_send_hint_budget(_max_send_in_flight_memory / max_hints_send_queue_length)
, _send_limiter(_max_send_in_flight_memory)
, _space_watchdog(_shard_managers, _per_device_limits_map)
{}
resource_manager(resource_manager&&) = delete;
resource_manager& operator=(resource_manager&&) = delete;
future<semaphore_units<semaphore_default_exception_factory>> get_send_units_for(size_t buf_size);
future<> start(shared_ptr<service::storage_proxy> proxy_ptr, shared_ptr<gms::gossiper> gossiper_ptr, shared_ptr<service::storage_service> ss_ptr);
void allow_replaying() noexcept;
future<> stop() noexcept;
void register_manager(manager& m);
future<> prepare_per_device_limits();
};
}
}