fix: Introduce retry for transient errors from DB (#3167)

This commit is contained in:
Alex Kremer
2026-08-10 14:39:15 +01:00
committed by GitHub
parent f490117cfb
commit dbdbe4ea1f
27 changed files with 665 additions and 88 deletions

View File

@@ -4,6 +4,7 @@
#include "data/LedgerCacheInterface.hpp"
#include "data/Types.hpp"
#include "etl/CorruptionDetector.hpp"
#include "util/Retry.hpp"
#include "util/Spawn.hpp"
#include "util/log/Logger.hpp"
@@ -35,41 +36,118 @@
namespace data {
/**
* @brief Represents a database timeout error.
* @brief Represents a transient database error that the caller should retry.
*/
class DatabaseTimeout : public std::exception {
class DatabaseError : public std::exception {
std::string message_{"Transient database error. Please retry the request"};
public:
DatabaseError() = default;
/**
* @brief Construct with a description of the underlying failure.
*
* @param message What actually went wrong.
*/
explicit DatabaseError(std::string message) : message_{std::move(message)}
{
}
/**
* @return The error message as a C string
*/
[[nodiscard]] char const*
what() const throw() override
what() const noexcept override
{
return "Database read timed out. Please retry the request";
return message_.c_str();
}
};
static constexpr std::size_t kDefaultWaitBetweenRetry = 500;
/**
* @brief A helper function that catches DatabaseTimeout exceptions and retries indefinitely.
* @brief Delay before the first retry in @ref retryOnTimeout().
*/
static constexpr std::chrono::milliseconds kDefaultWaitBetweenRetry{500};
/**
* @brief Default upper bound for the exponential backoff in @ref retryOnTimeout().
*/
static constexpr std::chrono::milliseconds kMaxWaitBetweenRetry{5'000};
/**
* @brief Default delays for @ref retryOnTimeout().
*/
static constexpr util::Retry::Delays kDefaultRetryDelays{
.initial = kDefaultWaitBetweenRetry,
.max = kDefaultWaitBetweenRetry
};
/**
* @brief Retry `func` while it throws DatabaseError, suspending the calling coroutine in between.
*
* @tparam FnType The type of function object to execute
* @param func The function object to execute
* @param waitMs Delay between retry attempts
* @param yield The coroutine to suspend between attempts
* @param delays The delays to use between attempts
* @return The same as the return type of func
*/
template <typename FnType>
auto
retryOnTimeout(FnType func, size_t waitMs = kDefaultWaitBetweenRetry)
retryOnTimeout(
FnType func,
boost::asio::yield_context yield,
util::Retry::Delays delays = kDefaultRetryDelays
)
{
static util::Logger const log{"Backend"}; // NOLINT(readability-identifier-naming)
auto retry = util::makeRetryExponentialBackoff(delays, yield.get_executor());
while (true) {
try {
return func();
} catch (DatabaseTimeout const&) {
LOG(log.error()) << "Database request timed out. Sleeping and retrying ... ";
std::this_thread::sleep_for(std::chrono::milliseconds(waitMs));
} catch (DatabaseError const& e) {
auto const delayMs =
std::chrono::duration_cast<std::chrono::milliseconds>(retry.delayValue()).count();
LOG(log.error()) << e.what() << " (attempt " << retry.attemptNumber() + 1
<< "). Retrying in " << delayMs << "ms ...";
retry.wait(yield);
}
}
}
/**
* @brief Retry `func` while it throws DatabaseError, blocking the calling thread in between.
*
* @warning Blocks the calling thread; from a coroutine use the `yield_context` overload instead.
*
* @tparam FnType The type of function object to execute
* @param func The function object to execute
* @param delays The delays to use between attempts
* @return The same as the return type of func
*/
template <typename FnType>
auto
retryOnTimeout(FnType func, util::Retry::Delays delays = kDefaultRetryDelays)
{
static util::Logger const log{"Backend"}; // NOLINT(readability-identifier-naming)
util::ExponentialBackoffStrategy backoff{delays};
std::size_t attempt = 1;
while (true) {
try {
return func();
} catch (DatabaseError const& e) {
auto const delay = backoff.getDelay();
LOG(log.error()) << e.what() << " (attempt " << attempt << "). Retrying in "
<< std::chrono::duration_cast<std::chrono::milliseconds>(delay).count()
<< "ms ...";
++attempt;
std::this_thread::sleep_for(delay);
backoff.increaseDelay();
}
}
}
@@ -105,18 +183,21 @@ synchronous(FnType&& func)
}
/**
* @brief Synchronously execute the given function object and retry until no DatabaseTimeout is
* @brief Synchronously execute the given function object and retry until no DatabaseError is
* thrown.
*
* @warning Blocks the calling thread while backing off.
*
* @tparam FnType The type of function object to execute
* @param func The function object to execute
* @param delays The delays to use between attempts
* @return The same as the return type of func
*/
template <typename FnType>
auto
synchronousAndRetryOnTimeout(FnType&& func)
synchronousAndRetryOnTimeout(FnType&& func, util::Retry::Delays delays = kDefaultRetryDelays)
{
return retryOnTimeout([&]() { return synchronous(func); });
return retryOnTimeout([&]() { return synchronous(func); }, delays);
}
/**
@@ -139,8 +220,27 @@ public:
BackendInterface(LedgerCacheInterface& cache) : cache_{cache}
{
}
virtual ~BackendInterface() = default;
/**
* @return Delay before the first retry of a request against this backend
*/
[[nodiscard]] virtual std::chrono::milliseconds
initialRetryDelay() const
{
return kDefaultWaitBetweenRetry;
}
/**
* @return Upper bound for the retry backoff; equal to @ref initialRetryDelay() means flat
*/
[[nodiscard]] virtual std::chrono::milliseconds
maxRetryDelay() const
{
return kMaxWaitBetweenRetry;
}
// TODO https://github.com/XRPLF/clio/issues/1956: Remove this hack once old ETL is removed.
// Cache should not be exposed thru BackendInterface
@@ -705,7 +805,7 @@ public:
hardFetchLedgerRange(boost::asio::yield_context yield) const = 0;
/**
* @brief Fetches the ledger range from DB retrying until no DatabaseTimeout is thrown.
* @brief Fetches the ledger range from DB retrying until no DatabaseError is thrown.
*
* @return The ledger range if available; nullopt otherwise
*/

View File

@@ -139,6 +139,24 @@ public:
*/
CassandraBackendFamily(CassandraBackendFamily&&) = delete;
/**
* @return The configured delay before the first retry
*/
[[nodiscard]] std::chrono::milliseconds
initialRetryDelay() const override
{
return settingsProvider_.getInitialRetryDelay();
}
/**
* @return The configured upper bound for the retry backoff
*/
[[nodiscard]] std::chrono::milliseconds
maxRetryDelay() const override
{
return settingsProvider_.getMaxRetryDelay();
}
TransactionsAndCursor
fetchAccountTransactions(
xrpl::AccountID const& account,

View File

@@ -80,17 +80,6 @@ public:
return code_;
}
/**
* @return true if the wrapped error is considered a timeout; false otherwise
*/
[[nodiscard]] bool
isTimeout() const
{
return code_ == CASS_ERROR_LIB_NO_HOSTS_AVAILABLE or
code_ == CASS_ERROR_LIB_REQUEST_TIMED_OUT or code_ == CASS_ERROR_SERVER_UNAVAILABLE or
code_ == CASS_ERROR_SERVER_OVERLOADED or code_ == CASS_ERROR_SERVER_READ_TIMEOUT;
}
/**
* @return true if the wrapped error is an invalid query; false otherwise
*/

View File

@@ -3,6 +3,7 @@
#include "data/cassandra/Types.hpp"
#include "data/cassandra/impl/Cluster.hpp"
#include "util/Constants.hpp"
#include "util/config/ConfigDefinition.hpp"
#include "util/config/ObjectView.hpp"
#include <cerrno>
@@ -25,6 +26,12 @@ SettingsProvider::SettingsProvider(util::config::ObjectView const& cfg)
, keyspace_{cfg.get<std::string>("keyspace")}
, tablePrefix_{cfg.maybeValue<std::string>("table_prefix")}
, replicationFactor_{cfg.get<uint16_t>("replication_factor")}
, initialRetryDelay_{util::config::ClioConfigDefinition::toMilliseconds(
cfg.get<float>("initial_request_retry_delay")
)}
, maxRetryDelay_{util::config::ClioConfigDefinition::toMilliseconds(
cfg.get<float>("max_request_retry_delay")
)}
, settings_{parseSettings()}
{
}
@@ -94,9 +101,9 @@ SettingsProvider::parseSettings() const
}
if (config_.getValueView("request_timeout").hasValue()) {
auto const requestTimeoutSecond = config_.get<uint32_t>("request_timeout");
settings.requestTimeout =
std::chrono::milliseconds{requestTimeoutSecond * util::kMillisecondsPerSecond};
settings.requestTimeout = util::config::ClioConfigDefinition::toMilliseconds(
config_.get<float>("request_timeout")
);
}
settings.certificate = parseOptionalCertificate();

View File

@@ -4,6 +4,7 @@
#include "data/cassandra/impl/Cluster.hpp"
#include "util/config/ObjectView.hpp"
#include <chrono>
#include <cstdint>
#include <optional>
#include <string>
@@ -19,6 +20,8 @@ class SettingsProvider {
std::string keyspace_;
std::optional<std::string> tablePrefix_;
uint16_t replicationFactor_;
std::chrono::milliseconds initialRetryDelay_;
std::chrono::milliseconds maxRetryDelay_;
Settings settings_;
public:
@@ -62,6 +65,24 @@ public:
return replicationFactor_;
}
/**
* @return Delay before the first retry of a failed request
*/
[[nodiscard]] std::chrono::milliseconds
getInitialRetryDelay() const
{
return initialRetryDelay_;
}
/**
* @return Upper bound for the retry backoff
*/
[[nodiscard]] std::chrono::milliseconds
getMaxRetryDelay() const
{
return maxRetryDelay_;
}
private:
[[nodiscard]] std::optional<std::string>
parseOptionalCertificate() const;

View File

@@ -15,6 +15,7 @@
#include <boost/asio/io_context.hpp>
#include <boost/asio/spawn.hpp>
#include <boost/json/object.hpp>
#include <fmt/format.h>
#include <algorithm>
#include <atomic>
@@ -27,6 +28,7 @@
#include <mutex>
#include <optional>
#include <stdexcept>
#include <string>
#include <thread>
#include <type_traits>
#include <vector>
@@ -169,7 +171,7 @@ public:
*
* @param preparedStatement Statement to prepare and execute
* @param args Args to bind to the prepared statement
* @throws DatabaseTimeout on timeout
* @throws DatabaseError on a database error
*/
template <typename... Args>
void
@@ -185,7 +187,7 @@ public:
* Retries forever with retry policy specified by @ref AsyncExecutor
*
* @param statement Statement to execute
* @throws DatabaseTimeout on timeout
* @throws DatabaseError on a database error
*/
void
write(StatementType&& statement)
@@ -215,7 +217,7 @@ public:
* Retries forever with retry policy specified by @ref AsyncExecutor.
*
* @param statements Vector of statements to execute as a batch
* @throws DatabaseTimeout on timeout
* @throws DatabaseError on a database error
*/
void
write(std::vector<StatementType>&& statements)
@@ -254,7 +256,7 @@ public:
* Retries forever with retry policy specified by @ref AsyncExecutor.
*
* @param statements Vector of statements to execute
* @throws DatabaseTimeout on timeout
* @throws DatabaseError on a database error
*/
void
writeEach(std::vector<StatementType>&& statements)
@@ -272,7 +274,7 @@ public:
* @param token Completion token (yield_context)
* @param preparedStatement Statement to prepare and execute
* @param args Args to bind to the prepared statement
* @throws DatabaseTimeout on timeout
* @throws DatabaseError on a database error
* @return ResultType or error wrapped in Expected
*/
template <typename... Args>
@@ -289,7 +291,7 @@ public:
*
* @param token Completion token (yield_context)
* @param statements Statements to execute in a batch
* @throws DatabaseTimeout on timeout
* @throws DatabaseError on a database error
* @return ResultType or error wrapped in Expected
*/
[[maybe_unused]] ResultOrErrorType
@@ -346,7 +348,7 @@ public:
*
* @param token Completion token (yield_context)
* @param statement Statement to execute
* @throws DatabaseTimeout on timeout
* @throws DatabaseError on a database error
* @return ResultType or error wrapped in Expected
*/
[[maybe_unused]] ResultOrErrorType
@@ -402,7 +404,7 @@ public:
*
* @param token Completion token (yield_context)
* @param statements Statements to execute
* @throws DatabaseTimeout on db error
* @throws DatabaseError on a database error
* @return Vector of results
*/
std::vector<ResultType>
@@ -457,7 +459,7 @@ public:
);
counters_->registerReadError(errorsCount);
counters_->registerReadFinished(startTime, statements.size() - errorsCount);
throw DatabaseTimeout{};
throw DatabaseError{};
}
counters_->registerReadFinished(startTime, statements.size());
@@ -551,11 +553,13 @@ private:
void
throwErrorIfNeeded(CassandraError err) const
{
if (err.isTimeout())
throw DatabaseTimeout();
// NOTE: etl::impl::Loader and etl::impl::Extractor treat std::runtime_error as
// "amendment blocked", so only genuinely permanent failures may be thrown as one.
if (err.isInvalidQuery())
throw std::runtime_error("Invalid query");
// anything else, including unclassified codes, is transient and gets retried
throw DatabaseError{fmt::format("Database error [{}]: {}", err.code(), err.message())};
}
};

View File

@@ -28,8 +28,7 @@ public:
ExponentialBackoffRetryPolicy(boost::asio::io_context& ioc)
: retry_(
util::makeRetryExponentialBackoff(
std::chrono::milliseconds(1),
std::chrono::seconds(1),
{.initial = std::chrono::milliseconds(1), .max = std::chrono::seconds(1)},
boost::asio::make_strand(ioc)
)
)

View File

@@ -145,9 +145,17 @@ private:
LOG(log_.debug()) << "Starting a cursor: " << xrpl::strHex(start);
while (not token.isStopRequested() and not cache_.get().isDisabled()) {
auto res = data::retryOnTimeout([this, seq, cachePageFetchSize, &start, token]() {
return backend_->fetchLedgerPage(start, seq, cachePageFetchSize, false, token);
});
auto res = data::retryOnTimeout(
[this, seq, cachePageFetchSize, &start, token]() {
return backend_->fetchLedgerPage(
start, seq, cachePageFetchSize, false, token
);
},
token,
util::Retry::Delays{
.initial = backend_->initialRetryDelay(), .max = backend_->maxRetryDelay()
}
);
cache_.get().update(res.objects, seq, true);

View File

@@ -58,7 +58,9 @@ SubscriptionSource::SubscriptionSource(
, subscriptions_(std::move(subscriptions))
, strand_(boost::asio::make_strand(ioContext))
, wsTimeout_(wsTimeout)
, retry_(util::makeRetryExponentialBackoff(retryDelay, kRetryMaxDelay, strand_))
, retry_(
util::makeRetryExponentialBackoff({.initial = retryDelay, .max = kMaxRetryDelay}, strand_)
)
, onConnect_(std::move(onConnect))
, onDisconnect_(std::move(onDisconnect))
, onLedgerClosed_(std::move(onLedgerClosed))

View File

@@ -75,7 +75,7 @@ private:
util::StopHelper stopHelper_;
static constexpr std::chrono::seconds kWsTimeout{30};
static constexpr std::chrono::seconds kRetryMaxDelay{30};
static constexpr std::chrono::seconds kMaxRetryDelay{30};
static constexpr std::chrono::seconds kRetryDelay{1};
public:

View File

@@ -239,7 +239,7 @@ public:
* @return The error message
*/
[[nodiscard]] char const*
what() const throw() override
what() const noexcept override
{
return msg_.c_str();
}
@@ -267,7 +267,7 @@ public:
* @return The error message
*/
[[nodiscard]] char const*
what() const throw() override
what() const noexcept override
{
return account_.c_str();
}

View File

@@ -178,8 +178,8 @@ public:
}
return Result{std::move(v)};
} catch (data::DatabaseTimeout const& t) {
LOG(log_.error()) << "Database timeout";
} catch (data::DatabaseError const& t) {
LOG(log_.error()) << "Database error: " << t.what();
notifyTooBusy();
return Result{Status{RippledError::RpcTooBusy}};
@@ -361,8 +361,8 @@ private:
}
return Result{std::move(v)};
} catch (data::DatabaseTimeout const& t) {
LOG(log_.error()) << "Database timeout";
} catch (data::DatabaseError const& t) {
LOG(log_.error()) << "Database error: " << t.what();
notifyTooBusy();
return Result{Status{RippledError::RpcTooBusy}};

View File

@@ -1,6 +1,8 @@
#include "util/Retry.hpp"
#include <boost/asio/any_io_executor.hpp>
#include <boost/asio/io_context.hpp>
#include <boost/asio/spawn.hpp>
#include <boost/asio/strand.hpp>
#include <algorithm>
@@ -42,6 +44,11 @@ Retry::Retry(
{
}
Retry::Retry(RetryStrategyPtr strategy, boost::asio::any_io_executor executor)
: strategy_(std::move(strategy)), timer_(executor)
{
}
Retry::~Retry()
{
*canceled_ = true;
@@ -73,11 +80,8 @@ Retry::reset()
(*strategy_).reset();
}
ExponentialBackoffStrategy::ExponentialBackoffStrategy(
std::chrono::steady_clock::duration delay,
std::chrono::steady_clock::duration maxDelay
)
: RetryStrategy(delay), maxDelay_(maxDelay)
ExponentialBackoffStrategy::ExponentialBackoffStrategy(Retry::Delays delays)
: RetryStrategy(delays.initial), maxDelay_(delays.max)
{
}
@@ -88,14 +92,32 @@ ExponentialBackoffStrategy::nextDelay() const
return std::min(next, maxDelay_);
}
void
Retry::wait(boost::asio::yield_context yield)
{
*canceled_ = false;
timer_.expires_after(strategy_->getDelay());
strategy_->increaseDelay();
++attemptNumber_;
// error ignored on purpose: a cancelled timer just means the caller retries sooner
boost::system::error_code ec;
timer_.async_wait(yield[ec]);
}
Retry
makeRetryExponentialBackoff(
std::chrono::steady_clock::duration delay,
std::chrono::steady_clock::duration maxDelay,
Retry::Delays delays,
boost::asio::strand<boost::asio::io_context::executor_type> strand
)
{
return Retry(std::make_unique<ExponentialBackoffStrategy>(delay, maxDelay), std::move(strand));
return Retry(std::make_unique<ExponentialBackoffStrategy>(delays), std::move(strand));
}
Retry
makeRetryExponentialBackoff(Retry::Delays delays, boost::asio::any_io_executor executor)
{
return Retry(std::make_unique<ExponentialBackoffStrategy>(delays), std::move(executor));
}
} // namespace util

View File

@@ -1,7 +1,9 @@
#pragma once
#include <boost/asio/any_io_executor.hpp>
#include <boost/asio/error.hpp>
#include <boost/asio/io_context.hpp>
#include <boost/asio/spawn.hpp>
#include <boost/asio/steady_timer.hpp>
#include <boost/asio/strand.hpp>
@@ -65,6 +67,16 @@ class Retry {
std::shared_ptr<std::atomic_bool> canceled_{std::make_shared<std::atomic_bool>(false)};
public:
/**
* @brief The delays to use between retry attempts
*
* Equal `initial` and `max` mean a flat delay with no backoff.
*/
struct Delays {
std::chrono::steady_clock::duration initial;
std::chrono::steady_clock::duration max;
};
/**
* @brief Construct a new Retry object
*
@@ -76,6 +88,16 @@ public:
boost::asio::strand<boost::asio::io_context::executor_type> strand
);
/**
* @brief Construct a new Retry object from any I/O executor
*
* For coroutines, pass `yield.get_executor()` and drive it with @ref wait().
*
* @param strategy The retry strategy to use
* @param executor The executor to run the retry timer on
*/
Retry(RetryStrategyPtr strategy, boost::asio::any_io_executor executor);
/**
* @brief Destroy the Retry object
*/
@@ -105,6 +127,17 @@ public:
});
}
/**
* @brief Wait out the current delay by suspending the calling coroutine, then back off.
*
* Unlike @ref retry() this returns once the delay elapsed instead of scheduling a callback, so
* the caller can keep its own loop. Advances the delay and attempt number like @ref retry().
*
* @param yield The coroutine to suspend
*/
void
wait(boost::asio::yield_context yield);
/**
* @brief Cancel scheduled retry if any
*/
@@ -140,13 +173,9 @@ public:
/**
* @brief Construct a new Exponential Backoff Strategy object
*
* @param delay The initial delay value
* @param maxDelay The maximum delay value
* @param delays The delays to use between attempts
*/
ExponentialBackoffStrategy(
std::chrono::steady_clock::duration delay,
std::chrono::steady_clock::duration maxDelay
);
explicit ExponentialBackoffStrategy(Retry::Delays delays);
private:
[[nodiscard]] std::chrono::steady_clock::duration
@@ -156,16 +185,24 @@ private:
/**
* @brief Create a retry mechanism with exponential backoff strategy
*
* @param delay The initial delay value
* @param maxDelay The maximum delay value
* @param delays The delays to use between attempts
* @param strand The strand to use for async operations
* @return The retry object
*/
Retry
makeRetryExponentialBackoff(
std::chrono::steady_clock::duration delay,
std::chrono::steady_clock::duration maxDelay,
Retry::Delays delays,
boost::asio::strand<boost::asio::io_context::executor_type> strand
);
/**
* @brief Create a retry mechanism with exponential backoff strategy on any I/O executor
*
* @param delays The delays to use between attempts
* @param executor The executor to run the retry timer on
* @return The retry object
*/
Retry
makeRetryExponentialBackoff(Retry::Delays delays, boost::asio::any_io_executor executor);
} // namespace util

View File

@@ -278,7 +278,15 @@ getClioConfig()
{"database.cassandra.connect_timeout",
ConfigValue{ConfigType::Integer}.optional().withConstraint(gValidateUint32)},
{"database.cassandra.request_timeout",
ConfigValue{ConfigType::Integer}.optional().withConstraint(gValidateUint32)},
ConfigValue{ConfigType::Double}.optional().withConstraint(gValidatePositiveDouble)},
{"database.cassandra.initial_request_retry_delay",
ConfigValue{ConfigType::Double}.defaultValue(0.5).withConstraint(
gValidatePositiveDouble
)},
{"database.cassandra.max_request_retry_delay",
ConfigValue{ConfigType::Double}.defaultValue(5.0).withConstraint(
gValidatePositiveDouble
)},
{"database.cassandra.username", ConfigValue{ConfigType::String}.optional()},
{"database.cassandra.password", ConfigValue{ConfigType::String}.optional()},
{"database.cassandra.certfile", ConfigValue{ConfigType::String}.optional()},

View File

@@ -182,8 +182,14 @@ This document provides a list of all available Clio configuration properties in
"established."},
KV{.key = "database.cassandra.request_timeout",
.value = "The maximum amount of time in seconds that the system waits for a request to "
"be fetched from the "
"database."},
"be fetched from the database. Should be set higher than the server side read "
"timeout. If omitted, no request timeout is applied."},
KV{.key = "database.cassandra.initial_request_retry_delay",
.value = "How long in seconds to wait before the first retry of a database request "
"that failed with a transient error."},
KV{.key = "database.cassandra.max_request_retry_delay",
.value = "Upper bound in seconds for the exponential backoff between retries of a "
"database request."},
KV{.key = "database.cassandra.username",
.value = "The username used for authenticating with the database."},
KV{.key = "database.cassandra.password",