Optimize apply speed under heavy write pressure (pingcap#4883)

ref pingcap#4728

dbms/src/Storages/DeltaMerge/Delta/DeltaValueSpace.cpp

@@ -141,6 +141,19 @@ bool DeltaValueSpace::ingestColumnFiles(DMContext & /*context*/, const RowKeyRan
 
 bool DeltaValueSpace::flush(DMContext & context)
 {
+    bool v = false;
+    if (!is_flushing.compare_exchange_strong(v, true))
+    {
+        // other thread is flushing, just return.
+        LOG_FMT_DEBUG(log, "{}, Flush stop because other thread is flushing", simpleInfo());
+        return false;
+    }
+    SCOPE_EXIT({
+        bool v = true;
+        if (!is_flushing.compare_exchange_strong(v, false))
+            throw Exception(simpleInfo() + " is expected to be flushing", ErrorCodes::LOGICAL_ERROR);
+    });
+
     LOG_FMT_DEBUG(log, "{}, Flush start", info());
 
     /// We have two types of data needed to flush to disk:

dbms/src/Storages/DeltaMerge/Delta/DeltaValueSpace.h

@@ -77,6 +77,11 @@ class DeltaValueSpace
     /// Note that those things can not be done at the same time.
     std::atomic_bool is_updating = false;
 
+    /// Note that it's safe to do multiple flush concurrently but only one of them can succeed,
+    /// and other thread's work is just a waste of resource.
+    /// So we only allow one flush task running at any time to aviod waste resource.
+    std::atomic_bool is_flushing = false;
+
     std::atomic<size_t> last_try_flush_rows = 0;
     std::atomic<size_t> last_try_flush_bytes = 0;
     std::atomic<size_t> last_try_compact_column_files = 0;
@@ -159,6 +164,8 @@ class DeltaValueSpace
     size_t getTotalCacheBytes() const;
     size_t getValidCacheRows() const;
 
+    bool isFlushing() const { return is_flushing; }
+
     bool isUpdating() const { return is_updating; }
 
     bool tryLockUpdating()

dbms/src/Storages/DeltaMerge/DeltaMergeStore.cpp

@@ -980,14 +980,14 @@ void DeltaMergeStore::deleteRange(const Context & db_context, const DB::Settings
         checkSegmentUpdate(dm_context, segment, ThreadType::Write);
 }
 
-void DeltaMergeStore::flushCache(const DMContextPtr & dm_context, const RowKeyRange & range)
+bool DeltaMergeStore::flushCache(const DMContextPtr & dm_context, const RowKeyRange & range, bool try_until_succeed)
 {
     RowKeyRange cur_range = range;
     while (!cur_range.none())
     {
         RowKeyRange segment_range;
 
-        // Keep trying until succeeded.
+        // Keep trying until succeeded if needed.
         while (true)
         {
             SegmentPtr segment;
@@ -1010,10 +1010,15 @@ void DeltaMergeStore::flushCache(const DMContextPtr & dm_context, const RowKeyRa
             {
                 break;
             }
+            else if (!try_until_succeed)
+            {
+                return false;
+            }
         }
 
         cur_range.setStart(segment_range.end);
     }
+    return true;
 }
 
 void DeltaMergeStore::mergeDeltaAll(const Context & context)
@@ -1347,6 +1352,12 @@ void DeltaMergeStore::checkSegmentUpdate(const DMContextPtr & dm_context, const
         && (delta_rows - delta_last_try_flush_rows >= delta_cache_limit_rows
             || delta_bytes - delta_last_try_flush_bytes >= delta_cache_limit_bytes);
     bool should_foreground_flush = unsaved_rows >= delta_cache_limit_rows * 3 || unsaved_bytes >= delta_cache_limit_bytes * 3;
+    /// For write thread, we want to avoid foreground flush to block the process of apply raft command.
+    /// So we increase the threshold of foreground flush for write thread.
+    if (thread_type == ThreadType::Write)
+    {
+        should_foreground_flush = unsaved_rows >= delta_cache_limit_rows * 10 || unsaved_bytes >= delta_cache_limit_bytes * 10;
+    }
 
     bool should_background_merge_delta = ((delta_check_rows >= delta_limit_rows || delta_check_bytes >= delta_limit_bytes) //
                                           && (delta_rows - delta_last_try_merge_delta_rows >= delta_cache_limit_rows
@@ -1404,9 +1415,16 @@ void DeltaMergeStore::checkSegmentUpdate(const DMContextPtr & dm_context, const
         }
         else if (should_background_flush)
         {
-            delta_last_try_flush_rows = delta_rows;
-            delta_last_try_flush_bytes = delta_bytes;
-            try_add_background_task(BackgroundTask{TaskType::Flush, dm_context, segment, {}});
+            /// It's meaningless to add more flush tasks if the segment is flushing.
+            /// Because only one flush task can proceed at any time.
+            /// And after the current flush task finished,
+            /// it will call `checkSegmentUpdate` again to check whether there is more flush task to do.
+            if (!segment->isFlushing())
+            {
+                delta_last_try_flush_rows = delta_rows;
+                delta_last_try_flush_bytes = delta_bytes;
+                try_add_background_task(BackgroundTask{TaskType::Flush, dm_context, segment, {}});
+            }
         }
     }
 
@@ -1502,7 +1520,12 @@ void DeltaMergeStore::checkSegmentUpdate(const DMContextPtr & dm_context, const
         return false;
     };
     auto try_bg_compact = [&]() {
-        if (should_compact)
+        /// Compact task should be a really low priority task.
+        /// And if the segment is flushing,
+        /// we should avoid adding background compact task to reduce lock contention on the segment and save disk throughput.
+        /// And after the current flush task complete,
+        /// it will call `checkSegmentUpdate` again to check whether there is other kinds of task to do.
+        if (should_compact && !segment->isFlushing())
         {
             delta_last_try_compact_column_files = column_file_count;
             try_add_background_task(BackgroundTask{TaskType::Compact, dm_context, segment, {}});

dbms/src/Storages/DeltaMerge/DeltaMergeStore.h

@@ -367,14 +367,14 @@ class DeltaMergeStore : private boost::noncopyable
                            const SegmentIdSet & read_segments = {},
                            size_t extra_table_id_index = InvalidColumnID);
 
-    /// Force flush all data to disk.
-    void flushCache(const Context & context, const RowKeyRange & range)
+    /// Try flush all data in `range` to disk and return whether the task succeed.
+    bool flushCache(const Context & context, const RowKeyRange & range, bool try_until_succeed = true)
     {
         auto dm_context = newDMContext(context, context.getSettingsRef());
-        flushCache(dm_context, range);
+        return flushCache(dm_context, range, try_until_succeed);
     }
 
-    void flushCache(const DMContextPtr & dm_context, const RowKeyRange & range);
+    bool flushCache(const DMContextPtr & dm_context, const RowKeyRange & range, bool try_until_succeed = true);
 
     /// Merge delta into the stable layer for all segments.
     ///

dbms/src/Storages/DeltaMerge/Segment.h

@@ -300,6 +300,8 @@ class Segment : private boost::noncopyable
 
     void drop(const FileProviderPtr & file_provider, WriteBatches & wbs);
 
+    bool isFlushing() const { return delta->isFlushing(); }
+
     RowsAndBytes getRowsAndBytesInRange(
         DMContext & dm_context,
         const SegmentSnapshotPtr & segment_snap,

dbms/src/Storages/IManageableStorage.h

@@ -68,7 +68,7 @@ class IManageableStorage : public IStorage
 
     virtual void flushCache(const Context & /*context*/) {}
 
-    virtual void flushCache(const Context & /*context*/, const DM::RowKeyRange & /*range_to_flush*/) {}
+    virtual bool flushCache(const Context & /*context*/, const DM::RowKeyRange & /*range_to_flush*/, [[maybe_unused]] bool try_until_succeed = true) { return true; }
 
     virtual BlockInputStreamPtr status() { return {}; }
 

dbms/src/Storages/StorageDeltaMerge.cpp

@@ -775,12 +775,12 @@ void StorageDeltaMerge::checkStatus(const Context & context)
 
 void StorageDeltaMerge::flushCache(const Context & context)
 {
-    flushCache(context, DM::RowKeyRange::newAll(is_common_handle, rowkey_column_size));
+    flushCache(context, DM::RowKeyRange::newAll(is_common_handle, rowkey_column_size), /* try_until_succeed */ true);
 }
 
-void StorageDeltaMerge::flushCache(const Context & context, const DM::RowKeyRange & range_to_flush)
+bool StorageDeltaMerge::flushCache(const Context & context, const DM::RowKeyRange & range_to_flush, bool try_until_succeed)
 {
-    getAndMaybeInitStore()->flushCache(context, range_to_flush);
+    return getAndMaybeInitStore()->flushCache(context, range_to_flush, try_until_succeed);
 }
 
 void StorageDeltaMerge::mergeDelta(const Context & context)

dbms/src/Storages/StorageDeltaMerge.h

@@ -73,7 +73,7 @@ class StorageDeltaMerge
 
     void flushCache(const Context & context) override;
 
-    void flushCache(const Context & context, const DM::RowKeyRange & range_to_flush) override;
+    bool flushCache(const Context & context, const DM::RowKeyRange & range_to_flush, bool try_until_succeed) override;
 
     /// Merge delta into the stable layer for all segments.
     ///

dbms/src/Storages/Transaction/KVStore.cpp

@@ -129,7 +129,7 @@ void KVStore::traverseRegions(std::function<void(RegionID, const RegionPtr &)> &
         callback(region.first, region.second);
 }
 
-void KVStore::tryFlushRegionCacheInStorage(TMTContext & tmt, const Region & region, Poco::Logger * log)
+bool KVStore::tryFlushRegionCacheInStorage(TMTContext & tmt, const Region & region, Poco::Logger * log, bool try_until_succeed)
 {
     auto table_id = region.getMappedTableID();
     auto storage = tmt.getStorages().get(table_id);
@@ -139,7 +139,7 @@ void KVStore::tryFlushRegionCacheInStorage(TMTContext & tmt, const Region & regi
                         "tryFlushRegionCacheInStorage can not get table for region {} with table id {}, ignored",
                         region.toString(),
                         table_id);
-        return;
+        return true;
     }
 
     try
@@ -151,14 +151,15 @@ void KVStore::tryFlushRegionCacheInStorage(TMTContext & tmt, const Region & regi
             region.getRange()->getMappedTableID(),
             storage->isCommonHandle(),
             storage->getRowKeyColumnSize());
-        storage->flushCache(tmt.getContext(), rowkey_range);
+        return storage->flushCache(tmt.getContext(), rowkey_range, try_until_succeed);
     }
     catch (DB::Exception & e)
     {
         // We can ignore if storage is already dropped.
         if (e.code() != ErrorCodes::TABLE_IS_DROPPED)
             throw;
     }
+    return true;
 }
 
 void KVStore::tryPersist(RegionID region_id)
@@ -366,12 +367,12 @@ EngineStoreApplyRes KVStore::handleUselessAdminRaftCmd(
             if (rows >= region_compact_log_min_rows.load(std::memory_order_relaxed)
                 || size_bytes >= region_compact_log_min_bytes.load(std::memory_order_relaxed))
             {
-                // if rows or bytes more than threshold, flush cache and perist mem data.
+                // if rows or bytes more than threshold, try to flush cache and persist mem data.
                 return true;
             }
             else
             {
-                // if thhere is little data in mem, wait until time interval reached threshold.
+                // if there is little data in mem, wait until time interval reached threshold.
                 // use random period so that lots of regions will not be persisted at same time.
                 auto compact_log_period = std::rand() % region_compact_log_period.load(std::memory_order_relaxed); // NOLINT
                 return !(curr_region.lastCompactLogTime() + Seconds{compact_log_period} > Clock::now());
@@ -381,11 +382,17 @@ EngineStoreApplyRes KVStore::handleUselessAdminRaftCmd(
 
     if (check_sync_log())
     {
-        tryFlushRegionCacheInStorage(tmt, curr_region, log);
-        persistRegion(curr_region, region_task_lock, "compact raft log");
-        curr_region.markCompactLog();
-        curr_region.cleanApproxMemCacheInfo();
-        return EngineStoreApplyRes::Persist;
+        if (tryFlushRegionCacheInStorage(tmt, curr_region, log, /* try_until_succeed */ false))
+        {
+            persistRegion(curr_region, region_task_lock, "compact raft log");
+            curr_region.markCompactLog();
+            curr_region.cleanApproxMemCacheInfo();
+            return EngineStoreApplyRes::Persist;
+        }
+        else
+        {
+            return EngineStoreApplyRes::None;
+        }
     }
     return EngineStoreApplyRes::None;
 }

dbms/src/Storages/Transaction/KVStore.h

@@ -91,7 +91,7 @@ class KVStore final : private boost::noncopyable
 
     void tryPersist(RegionID region_id);
 
-    static void tryFlushRegionCacheInStorage(TMTContext & tmt, const Region & region, Poco::Logger * log);
+    static bool tryFlushRegionCacheInStorage(TMTContext & tmt, const Region & region, Poco::Logger * log, bool try_until_succeed = true);
 
     size_t regionSize() const;
     EngineStoreApplyRes handleAdminRaftCmd(raft_cmdpb::AdminRequest && request,

dbms/src/Storages/Transaction/RegionTable.cpp

@@ -230,7 +230,7 @@ void removeObsoleteDataInStorage(
         auto rowkey_range
             = DM::RowKeyRange::fromRegionRange(handle_range, table_id, table_id, storage->isCommonHandle(), storage->getRowKeyColumnSize());
         dm_storage->deleteRange(rowkey_range, context->getSettingsRef());
-        dm_storage->flushCache(*context, rowkey_range); // flush to disk
+        dm_storage->flushCache(*context, rowkey_range, /*try_until_succeed*/ true); // flush to disk
     }
     catch (DB::Exception & e)
     {