MDEV-11934: MariaRocks: Group Commit with binlog

Port the previous patch:
- Implement MariaDB's Group Commit API. This is a first
   attempt which lacks the expected performance.

To newer MariaDB (which includes newer MyRocks)

1 files changed, 99 insertions, 7 deletions

diff --git a/storage/rocksdb/ha_rocksdb.cc b/storage/rocksdb/ha_rocksdb.cc
index 17ea7b74565..3ca58b3f935 100644
--- a/storage/rocksdb/ha_rocksdb.cc
+++ b/storage/rocksdb/ha_rocksdb.cc
@@ -1807,6 +1807,12 @@ public:
   int64_t m_n_mysql_tables_in_use = 0;
 
   /*
+    MariaDB's group commit:
+  */
+  bool commit_ordered_done;
+  bool commit_ordered_res;
+
+  /*
     for distinction between rdb_transaction_impl and rdb_writebatch_impl
     when using walk tx list
   */
@@ -2431,6 +2437,8 @@ public:
         THDVAR(m_thd, write_ignore_missing_column_families);
     m_is_two_phase = rocksdb_enable_2pc;
 
+    commit_ordered_done= false;
+
     /*
       If m_rocksdb_reuse_tx is null this will create a new transaction object.
       Otherwise it will reuse the existing one.
@@ -2643,6 +2651,7 @@ public:
   bool is_tx_started() const override { return (m_batch != nullptr); }
 
   void start_tx() override {
+    commit_ordered_done= false; // Do we need this here?
     reset();
     write_opts.sync = (rocksdb_flush_log_at_trx_commit == FLUSH_LOG_SYNC);
     write_opts.disableWAL = THDVAR(m_thd, write_disable_wal);
@@ -2831,8 +2840,7 @@ static bool rocksdb_flush_wal(handlerton* hton __attribute__((__unused__)))
 */
 static int rocksdb_prepare(handlerton* hton, THD* thd, bool prepare_tx)
 {
-// This is "ASYNC_COMMIT" feature which is only in webscalesql
-  bool async=false;
+  bool async=false; // This is "ASYNC_COMMIT" feature which is only present in webscalesql
 
   Rdb_transaction *&tx = get_tx_from_thd(thd);
   if (!tx->can_prepare()) {
@@ -2842,7 +2850,8 @@ static int rocksdb_prepare(handlerton* hton, THD* thd, bool prepare_tx)
       (!my_core::thd_test_options(thd, OPTION_NOT_AUTOCOMMIT | OPTION_BEGIN))) {
     /* We were instructed to prepare the whole transaction, or
     this is an SQL statement end and autocommit is on */
-#ifdef MARIAROCKS_NOT_YET // disable prepare/commit
+
+#ifdef MARIAROCKS_NOT_YET // Crash-safe slave does not work yet
     std::vector<st_slave_gtid_info> slave_gtid_info;
     my_core::thd_slave_gtid_info(thd, &slave_gtid_info);
     for (const auto &it : slave_gtid_info) {
@@ -2852,22 +2861,41 @@ static int rocksdb_prepare(handlerton* hton, THD* thd, bool prepare_tx)
 #endif
 
     if (tx->is_two_phase()) {
+
+      /*
+        MariaDB: the following branch is never taken.
+        We always flush at Prepare and rely on RocksDB's internal Group Commit
+        to do some grouping.
+      */
       if (thd->durability_property == HA_IGNORE_DURABILITY || async) {
         tx->set_sync(false);
       }
+
+      /*
+        MariaDB: do not flush logs if we are running in a non-crash-safe mode.
+      */
+      if (!rocksdb_flush_log_at_trx_commit)
+        tx->set_sync(false);
+
       XID xid;
       thd_get_xid(thd, reinterpret_cast<MYSQL_XID *>(&xid));
       if (!tx->prepare(rdb_xid_to_string(xid))) {
         return HA_EXIT_FAILURE;
       }
-      if (thd->durability_property == HA_IGNORE_DURABILITY )
+
+      /*
+        MariaDB: our Group Commit implementation does not use the
+        hton->flush_logs call (at least currently) so the following is not
+        needed (TODO: will we need this for binlog rotation?)
+      */
 #ifdef MARIAROCKS_NOT_YET      
-          (rocksdb_flush_log_at_trx_commit != FLUSH_LOG_NEVER)) {
+      if (thd->durability_property == HA_IGNORE_DURABILITY )
+          (rocksdb_flush_log_at_trx_commit != FLUSH_LOG_NEVER))
           &&
           THDVAR(thd, flush_log_at_trx_commit)) 
 #endif          
-      {
 #ifdef MARIAROCKS_NOT_YET
+      {
         // MariaRocks: disable the
         //   "write/sync redo log before flushing binlog cache to file"
         //  feature. See a869c56d361bb44f46c0efeb11a8f03561676247
@@ -2875,8 +2903,8 @@ static int rocksdb_prepare(handlerton* hton, THD* thd, bool prepare_tx)
           we set the log sequence as '1' just to trigger hton->flush_logs
         */
         thd_store_lsn(thd, 1, DB_TYPE_ROCKSDB);
-#endif        
       }
+#endif
     }
 
     DEBUG_SYNC(thd, "rocksdb.prepared");
@@ -3026,6 +3054,50 @@ static int rocksdb_recover(handlerton* hton, XID* xid_list, uint len)
   return count;
 }
 
+
+/*
+  Handle a commit checkpoint request from server layer.
+
+  InnoDB does this:
+    We put the request in a queue, so that we can notify upper layer about
+    checkpoint complete when we have flushed the redo log.
+    If we have already flushed all relevant redo log, we notify immediately.
+
+  MariaRocks just flushes everything right away ATM
+*/
+
+static void rocksdb_checkpoint_request(handlerton *hton,
+                                       void *cookie)
+{
+  const rocksdb::Status s= rdb->SyncWAL();
+  //TODO: what to do on error?
+  if (s.ok())
+  {
+    rocksdb_wal_group_syncs++;
+    commit_checkpoint_notify_ha(hton, cookie);
+  }
+}
+
+/*
+  @param all:   TRUE - commit the transaction
+                FALSE - SQL statement ended
+*/
+static void rocksdb_commit_ordered(handlerton *hton, THD* thd, bool all)
+{
+  // Same assert as InnoDB has
+  DBUG_ASSERT(all || (!thd_test_options(thd, OPTION_NOT_AUTOCOMMIT |
+                                             OPTION_BEGIN)));
+  Rdb_transaction *&tx = get_tx_from_thd(thd);
+
+  tx->set_sync(false);
+
+  /* This will note the master position also */
+  tx->commit_ordered_res= tx->commit();
+  tx->commit_ordered_done= true;
+
+}
+
+
 static int rocksdb_commit(handlerton* hton, THD* thd, bool commit_tx)
 {
   DBUG_ENTER_FUNC();
@@ -3046,6 +3118,16 @@ static int rocksdb_commit(handlerton* hton, THD* thd, bool commit_tx)
     if (commit_tx || (!my_core::thd_test_options(thd, OPTION_NOT_AUTOCOMMIT |
                                                           OPTION_BEGIN))) {
       /*
+        This will not add anything to commit_latency_stats, and this is correct
+        right?
+      */
+      if (tx->commit_ordered_done)
+      {
+        thd_wakeup_subsequent_commits(thd, 0);
+        DBUG_RETURN((tx->commit_ordered_res? HA_ERR_INTERNAL_ERROR: 0));
+      }
+
+      /*
         We get here
          - For a COMMIT statement that finishes a multi-statement transaction
          - For a statement that has its own transaction
@@ -3053,6 +3135,7 @@ static int rocksdb_commit(handlerton* hton, THD* thd, bool commit_tx)
       if (tx->commit()) {
         DBUG_RETURN(HA_ERR_ROCKSDB_COMMIT_FAILED);
       }
+      thd_wakeup_subsequent_commits(thd, 0);
     } else {
       /*
         We get here when committing a statement within a transaction.
@@ -3076,6 +3159,7 @@ static int rocksdb_commit(handlerton* hton, THD* thd, bool commit_tx)
   DBUG_RETURN(HA_EXIT_SUCCESS);
 }
 
+
 static int rocksdb_rollback(handlerton *const hton, THD *const thd,
                             bool rollback_tx) {
   Rdb_transaction *&tx = get_tx_from_thd(thd);
@@ -3882,11 +3966,19 @@ static int rocksdb_init_func(void *const p) {
   rocksdb_hton->state = SHOW_OPTION_YES;
   rocksdb_hton->create = rocksdb_create_handler;
   rocksdb_hton->close_connection = rocksdb_close_connection;
+
   rocksdb_hton->prepare = rocksdb_prepare;
+  rocksdb_hton->prepare_ordered = NULL; // Do not need it
+
   rocksdb_hton->commit_by_xid = rocksdb_commit_by_xid;
   rocksdb_hton->rollback_by_xid = rocksdb_rollback_by_xid;
   rocksdb_hton->recover = rocksdb_recover;
+
+  rocksdb_hton->commit_ordered= rocksdb_commit_ordered;
   rocksdb_hton->commit = rocksdb_commit;
+
+  rocksdb_hton->commit_checkpoint_request= rocksdb_checkpoint_request;
+
   rocksdb_hton->rollback = rocksdb_rollback;
   rocksdb_hton->show_status = rocksdb_show_status;
   rocksdb_hton->start_consistent_snapshot =