OceanBase中的写入限速机制旨在控制系统中写入操作(一般写入操作包括插入、更新和删除等)的速率,目的是为了提高数据库系统的稳定性。本文主要通过以下2个参数来解释写入限速的实现机制。
当执行写入操作申请内存时,触发写入限速条件:已使用的 Memstore 内存超过设定的比例(比例阈值由 writing_throttling_trigger_percentage 参数确定),系统进入限速逻辑。
限速逻辑会将本次申请内存的任务分成多次进行限速。每次限速的执行时间最多为20毫秒。
如果上述任意一项条件满足,系统将退出限速循环,并将SQL完成的信息发送给客户端。这样可以确保SQL能够成功执行完成,并保证系统的稳定性。
以下通过源码以一条insert语句的部分堆栈来解释writing_throttling_trigger_percentage和writing_throttling_maximum_duration是如何影响限速逻辑的。
ObTablet::insert_row_without_rowkey_check()
int ObTablet::insert_row_without_rowkey_check(
ObRelativeTable &relative_table,
ObStoreCtx &store_ctx,
const common::ObIArray<share::schema::ObColDesc> &col_descs,
const storage::ObStoreRow &row)
{
int ret = OB_SUCCESS;
{
// insert_row_without_rowkey_check 执行结束时会调用ObStorageTableGuard的析构函数,进行限速处理
ObStorageTableGuard guard(this, store_ctx, true);
ObMemtable *write_memtable = nullptr;
...
//write_memtable->set()会调用ObFifoArena::alloc()申请内存, 在分配内存时进行限速判断
else if (OB_FAIL(write_memtable->set(store_ctx, relative_table.get_table_id(),
full_read_info_, col_descs, row)))
...
}
return ret;
}
该方法会实例化 ObStorageTableGuard类 , 限速的执行过程定义在该类的析构函数内, 所以程序会在执行完 write_memtable 后才进行限速。后续会进行写Memtable的流程,这里不做赘述, 大致调用堆栈如下:
| > oceanbase::storage::ObTablet::insert_row_without_rowkey_check(...) (/src/storage/tablet/ob_tablet.cpp:1425)
| + > oceanbase::memtable::ObMemtable::set(...) (/src/storage/memtable/ob_memtable.cpp:339)
| + - > oceanbase::memtable::ObMemtable::set_(...) (/src/storage/memtable/ob_memtable.cpp:2538)
| + - x > oceanbase::memtable::ObMemtable::mvcc_write_(...) (/src/storage/memtable/ob_memtable.cpp:2655)
| + - x = > oceanbase::memtable::ObMvccEngine::create_kv(...) (/src/storage/memtable/mvcc/ob_mvcc_engine.cpp:275)
| + - x = | > oceanbase::memtable::ObMTKVBuilder::dup_key(...) (/src/storage/memtable/ob_memtable.h:77)
| + - x = | + > oceanbase::common::ObGMemstoreAllocator::AllocHandle::alloc(...) (/src/share/allocator/ob_gmemstore_allocator.h:84)
| + - x = | + - > oceanbase::common::ObGMemstoreAllocator::alloc(...) (/src/share/allocator/ob_gmemstore_allocator.cpp:125)
| + - x = | + - x > oceanbase::common::ObFifoArena::alloc(...) (/src/share/allocator/ob_fifo_arena.cpp:157)
| + - x = | + - x = > oceanbase::common::ObFifoArena::speed_limit(...)(/src/share/allocator/ob_fifo_arena.cpp:301)
| + - x = | + - x = | > oceanbase::common::ObFifoArena::ObWriteThrottleInfo::check_and_calc_decay_factor(...)(/src/share/allocator/ob_fifo_arena.cpp:75)
| + > oceanbase::storage::ObStorageTableGuard::~ObStorageTableGuard(...) (/src/storage/ob_storage_table_guard.cpp:53)
ObFifoArena::alloc()
写memtable时会申请内存, 这时候会去判断是否需要限速
void* ObFifoArena::alloc(int64_t adv_idx, Handle& handle, int64_t size)
{
int ret = OB_SUCCESS;
void* ptr = NULL;
int64_t rsize = size + sizeof(Page) + sizeof(Ref);
// 调用speed limit 判断限速
speed_limit(ATOMIC_LOAD(&hold_), size);
...
}
ObFifoArena::speed_limit()
这个方法主要用来判断是否需要限速,同时根据配置的writing_throttling_maximum_duration值,计算出一个衰减因子用于等待时间的计算
void ObFifoArena::speed_limit(const int64_t cur_mem_hold, const int64_t alloc_size)
{
int ret = OB_SUCCESS;
//获取租户的writing_throttling_trigger_percentage值
int64_t trigger_percentage = get_writing_throttling_trigger_percentage_();
int64_t trigger_mem_limit = 0;
bool need_speed_limit = false;
int64_t seq = 0;
int64_t throttling_interval = 0;
// trigger_percentage <100 ,表示开启限速,再进行内存使用是否达到触发阈值的判断
if (trigger_percentage < 100) {
if (OB_UNLIKELY(cur_mem_hold < 0 || alloc_size <= 0 || lastest_memstore_threshold_ <= 0 || trigger_percentage <= 0)) {
COMMON_LOG(ERROR, "invalid arguments", K(cur_mem_hold), K(alloc_size), K(lastest_memstore_threshold_), K(trigger_percentage));
} else if (cur_mem_hold > (trigger_mem_limit = lastest_memstore_threshold_ * trigger_percentage / 100)) {
// 当前使用内存超过触发阈值,需要限速,设置need_speed_limit 为true
need_speed_limit = true;
// 获取writing_throttling_maximum_duration的值,默认 2h
int64_t alloc_duration = get_writing_throttling_maximum_duration_();
// 计算衰减因子,用于sleep时间计算
if (OB_FAIL(throttle_info_.check_and_calc_decay_factor(lastest_memstore_threshold_, trigger_percentage, alloc_duration))) {
COMMON_LOG(WARN, "failed to check_and_calc_decay_factor", K(cur_mem_hold), K(alloc_size), K(throttle_info_));
}
}
//这块代码是将内存和时钟值绑定,确保内存分配和写入限速的稳定性
advance_clock();
seq = ATOMIC_AAF(&max_seq_, alloc_size);
get_seq() = seq;
// 将need_speed_limit 赋值给tl_need_speed_limit 线程变量
tl_need_speed_limit() = need_speed_limit;
//日志记录,限速信息
if (need_speed_limit && REACH_TIME_INTERVAL(1 * 1000 * 1000L)) {
COMMON_LOG(INFO, "report write throttle info", K(alloc_size), K(attr_), K(throttling_interval),
"max_seq_", ATOMIC_LOAD(&max_seq_), K(clock_),
K(cur_mem_hold), K(throttle_info_), K(seq));
}
}
}
ObFifoArena::ObWriteThrottleInfo::check_and_calc_decay_factor()
计算衰减因子
int ObFifoArena::ObWriteThrottleInfo::check_and_calc_decay_factor(int64_t memstore_threshold,
int64_t trigger_percentage,
int64_t alloc_duration)
{
int ret = OB_SUCCESS;
if (memstore_threshold != memstore_threshold_
|| trigger_percentage != trigger_percentage_
|| alloc_duration != alloc_duration_
|| decay_factor_ <= 0) {
memstore_threshold_ = memstore_threshold;
trigger_percentage_ = trigger_percentage;
alloc_duration_ = alloc_duration;
int64_t available_mem = (100 - trigger_percentage_) * memstore_threshold_ / 100;
double N = static_cast<double>(available_mem) / static_cast<double>(MEM_SLICE_SIZE);
decay_factor_ = (static_cast<double>(alloc_duration) - N * static_cast<double>(MIN_INTERVAL))/ static_cast<double>((((N*(N+1)*N*(N+1)))/4));
decay_factor_ = decay_factor_ < 0 ? 0 : decay_factor_;
COMMON_LOG(INFO, "recalculate decay factor", K(memstore_threshold_), K(trigger_percentage_),
K(decay_factor_), K(alloc_duration), K(available_mem), K(N));
}
return ret;
}
decay_factor公式中,alloc_duration为writing_throttling_maximum_duration的值,4.0版本中为2h,MIN_INTERVAL默认值20ms。
简单来说,这个衰减因子是根据当前可用内存和writing_throttling_maximum_duration的值通过一个多项式计算出来的,整个过程如果writing_throttling_maximum_duration值不做调整,每次休眠时间会随着可用内存逐渐减少而慢慢增加。
ObStorageTableGuard::~ObStorageTableGuard()
限速流程执行
ObStorageTableGuard::~ObStorageTableGuard()
{
//tl_need_speed_limit 在ObFifoArena::alloc()方法中赋值
bool &need_speed_limit = tl_need_speed_limit();
// 在写操作的上下文中, 创建ObStorageTableGuard 实例时,need_control_mem_ 会被赋值为true
if (need_control_mem_ && need_speed_limit) {
bool need_sleep = true;
int64_t left_interval = SPEED_LIMIT_MAX_SLEEP_TIME;
//SPEED_LIMIT_MAX_SLEEP_TIME 默认20s,表示最大sleep时间
if (!for_replay_) {
// 如果不是回放日志
//store_ctx_.timeout_ - ObTimeUtility::current_time() 表示距离事务超时还要多久,如果该值小于0,表示事务已经超时
//两者取小
left_interval = min(SPEED_LIMIT_MAX_SLEEP_TIME, store_ctx_.timeout_ - ObTimeUtility::current_time());
}
// 如果memtable是冻结状态,不需要限速
if (NULL != memtable_) {
need_sleep = memtable_->is_active_memtable();
}
uint64_t timeout = 10000;//10s
//事件记录, 可以在v$session_event中查看,event名: memstore memory page alloc wait
//可以通过sql: select * from v$session_event where EVENT='memstore memory page alloc wait' 查询;
common::ObWaitEventGuard wait_guard(common::ObWaitEventIds::MEMSTORE_MEM_PAGE_ALLOC_WAIT, timeout, 0, 0, left_interval);
reset();
int tmp_ret = OB_SUCCESS;
bool has_sleep = false;
int64_t sleep_time = 0;
int time = 0;
int64_t &seq = get_seq();
if (store_ctx_.mvcc_acc_ctx_.is_write()) {
ObGMemstoreAllocator* memstore_allocator = NULL;
//获取当前租户的memstore内存分配器
if (OB_SUCCESS != (tmp_ret = ObMemstoreAllocatorMgr::get_instance().get_tenant_memstore_allocator(
MTL_ID(), memstore_allocator))) {
} else if (OB_ISNULL(memstore_allocator)) {
LOG_WARN_RET(OB_ALLOCATE_MEMORY_FAILED, "get_tenant_mutil_allocator failed", K(store_ctx_.tablet_id_), K(tmp_ret));
} else {
while (need_sleep &&
!memstore_allocator->check_clock_over_seq(seq) &&
(left_interval > 0)) {
if (for_replay_) {
// 如果是回放日志,并且当前租户下有正在进行的日志流,不做休眠,直接break
if(MTL(ObTenantFreezer *)->exist_ls_freezing()) {
break;
}
}
//计算休眠时间
int64_t expected_wait_time = memstore_allocator->expected_wait_time(seq);
if (expected_wait_time == 0) {
break;
}
//SLEEP_INTERVAL_PER_TIME 单次休眠时间,默认20ms
//线程休眠,每次最多20ms
uint32_t sleep_interval =
static_cast<uint32_t>(min(min(left_interval, SLEEP_INTERVAL_PER_TIME), expected_wait_time));
::usleep(sleep_interval);
// 累加休眠时间
sleep_time += sleep_interval;
// 休眠次数
time++;
//每次休眠之后,减去休眠时间
left_interval -= sleep_interval;
has_sleep = true;
//每次休眠之后,重新判断是否需要限速,因为可能在休眠过程中,内存经过转储后已经释放出来了,这时候就不需要继续限速了
need_sleep = memstore_allocator->need_do_writing_throttle();
}
}
}
// 日志记录,限速执行详情
if (REACH_TIME_INTERVAL(100 * 1000L) &&
sleep_time > 0) {
int64_t cost_time = ObTimeUtility::current_time() - init_ts_;
LOG_INFO("throttle situation", K(sleep_time), K(time), K(seq), K(for_replay_), K(cost_time));
}
if (for_replay_ && has_sleep) {
get_replay_is_writing_throttling() = true;
}
}
reset();
}
总结
OB的写入限速功能是在ObStorageTableGuard类的析构函数中实现的。由于该函数会在memtable写入完成后才被调用,因此限速行为是后置的,会影响下一次内存分配。换言之,在当前写入操作完成后,才会判断是否需要执行限速,若需要,会延迟下一次内存分配。这种设计既确保限速不会影响当前的写入操作,又能有效控制内存的分配和消耗。
该参数是租户级别的参数,可以在租户管理员账号下或者在sys租户中指定租户,设置内存写入达到 80% 开始限速,并保证剩余内存足够提供 2h 的写入限速,示例:
obclient> ALTER SYSTEM SET writing_throttling_trigger_percentage = 80;
Query OK, 0 rows affected
obclient> ALTER SYSTEM SET writing_throttling_maximum_duration = '2h';
Query OK, 0 rows affected
或者在sys租户中指定租户
obclient> ALTER SYSTEM SET writing_throttling_trigger_percentage = 80 tenant=<tenant_name>;
1.创建租户时使用 在写压力比较大的情况下,比如做导入数据时,限制写入速度也是一种简单高效的解决方法,虽然OceanBase的LSM-Tree存储引擎架构可以及时冻结memtable并释放内存,但在写入速度高于转储速度的场景下,仍有可能导致Memstore耗尽。最新版本4.0默认开启此配置,结合转储配置,可以有效控制Memstore的消耗。
2.发现qps异常下降时,尤其是包含大量写时,也可以通过以下方式确认是否是由于写入限制导致。
如果是触发限速导致的qps值下降,根据上面的代码分析可知,会记录在session_event表中,事件名是“memstore memory page alloc wait”。
select * from v$session_event where EVENT='memstore memory page alloc wait' \G;
*************************** 94. row ***************************
CON_ID: 1
SVR_IP: 10.186.64.124
SVR_PORT: 22882
SID: 3221487713
EVENT: memstore memory page alloc wait
TOTAL_WAITS: 182673
TOTAL_TIMEOUTS: 0
TIME_WAITED: 1004.4099
AVERAGE_WAIT: 0.005498403704981032
MAX_WAIT: 12.3022
TIME_WAITED_MICRO: 10044099
CPU: NULL
EVENT_ID: 11015
WAIT_CLASS_ID: 109
WAIT_CLASS#: 9
WAIT_CLASS: SYSTEM_IO
通过grep 'report write throttle info' observer.log ,如果输入如下日志就可以确定是由于限速导致的。
[2023-04-17 17:17:30.695621] INFO [COMMON] speed_limit (ob_fifo_arena.cpp:319) [26466][T1_L0_G0][T1][Y59620ABA407C-0005F9818D1BFE06-0-0] [lt=2] report write throttle info(alloc_size=32, attr_=tenant_id=1, label=Memstore, ctx_id=1, prio=0, throttling_interval=0, max_seq_=11045142952, clock_=11045143112, cur_mem_hold=524288000, throttle_info_={decay_factor_:"6.693207379708156213e-02", alloc_duration_:7200000000, trigger_percentage_:21, memstore_threshold_:2147483600, period_throttled_count_:0, period_throttled_time_:0, total_throttled_count_:0, total_throttled_time_:0}, seq=11045142952)
同时grep 'throttle situation' observer.log,可以看到这次限速的具体内容。
[2023-04-17 17:17:31.006880] INFO [STORAGE] ~ObStorageTableGuard (ob_storage_table_guard.cpp:109) [26466][T1_L0_G0][T1][Y59620ABA407C-0005F9818D1BFE06-0-0] [lt=85] throttle situation(sleep_time=4, time=1, seq=11048795064, for_replay_=false, cost_time=7025)
本文关键字: #Oceanbase# #写入限速#
原创声明:本文系作者授权腾讯云开发者社区发表,未经许可,不得转载。
如有侵权,请联系 cloudcommunity@tencent.com 删除。
原创声明:本文系作者授权腾讯云开发者社区发表,未经许可,不得转载。
如有侵权,请联系 cloudcommunity@tencent.com 删除。