revision-id: 4da9cb93fb8c291a0023451963c15fd2971b0210 (fb-prod201801-192-g4da9cb93fb8) parent(s): 70d97cc103fd98c7a4952e7b3a54f272fa7b36f4 author: Sergei Petrunia committer: Sergei Petrunia timestamp: 2019-01-21 21:36:09 +0300 message: Use InnoDB-like transaction isolation with Range Locking mode. DML statements (UPDATE/DELETE/..) will always read the latest committed data (as opposed to transaction's snapshot). --- mysql-test/suite/rocksdb/r/range_locking.result | 109 ++++++++++++++++++++++++ mysql-test/suite/rocksdb/t/range_locking.test | 95 +++++++++++++++++++++ storage/rocksdb/ha_rocksdb.cc | 81 +++++++++++++++--- 3 files changed, 271 insertions(+), 14 deletions(-) diff --git a/mysql-test/suite/rocksdb/r/range_locking.result b/mysql-test/suite/rocksdb/r/range_locking.result index a43f7d668d4..b0217d5269a 100644 --- a/mysql-test/suite/rocksdb/r/range_locking.result +++ b/mysql-test/suite/rocksdb/r/range_locking.result @@ -192,3 +192,112 @@ rollback; disconnect con1; connection default; drop table t0,t1; +# +# Transaction isolation test +# +create table t1 (pk int primary key, a int) engine=rocksdb; +insert into t1 values (1,1),(2,2),(3,3); +connect con1,localhost,root,,; +# TRX1: Start, Allocate a snapshot +connection con1; +begin; +select * from t1; +pk a +1 1 +2 2 +3 3 +# TRX2: Make a change that TRX1 will not see +connection default; +update t1 set a=2222 where pk=2; +# TRX1: Now, make a change that would overwrite TRX2'x change and commit +connection con1; +update t1 set a=a+1 where pk=2; +commit; +# Examine the result: +# pk=2, a=2223 means UPDATE in TRX1 used "read committed" (InnoDB-like isolation) +# pk=2, a=3 means UPDATE in TRX1 silently overwrote TRX2 +# (and with key tracking, one would get an error on the second UPDATE) +connection default; +select * from t1; +pk a +1 1 +2 2223 +3 3 +disconnect con1; +connection default; +drop table t1; +# +# Same test as above, but check the range scan +# +create table t1 (pk int primary key, a int) engine=rocksdb; +insert into t1 values (1,1),(2,2),(3,3),(4,4),(5,5),(6,6); +connect con1,localhost,root,,; +# TRX1: Start, Allocate a snapshot +connection con1; +begin; +select * from t1; +pk a +1 1 +2 2 +3 3 +4 4 +5 5 +6 6 +# TRX2: Make a change that TRX1 will not see +connection default; +update t1 set a=2222 where pk between 3 and 5; +# TRX1: Now, make a change that would overwrite TRX2'x change and commit +connection con1; +update t1 set a=a+1 where pk between 3 and 5; +commit; +# Examine the result: +# pk={3,4,5} a=2223 means UPDATE in TRX1 used "read committed" (InnoDB-like isolation) +connection default; +select * from t1; +pk a +1 1 +2 2 +3 2223 +4 2223 +5 2223 +6 6 +disconnect con1; +connection default; +drop table t1; +# +# Same as above, but test SELECT FOR UPDATE. +# +create table t1 (pk int primary key, a int) engine=rocksdb; +insert into t1 values (1,1),(2,2),(3,3),(4,4),(5,5),(6,6); +connect con1,localhost,root,,; +# TRX1: Start, Allocate a snapshot +connection con1; +begin; +select * from t1; +pk a +1 1 +2 2 +3 3 +4 4 +5 5 +6 6 +# TRX2: Make a change that TRX1 will not see +connection default; +update t1 set a=222 where pk=2; +update t1 set a=333 where pk=3; +# TRX1: Check what select [FOR UPDATE] sees +connection con1; +select * from t1 where pk in (2,3); +pk a +2 2 +3 3 +select * from t1 where pk=2 for update; +pk a +2 222 +select * from t1 where pk=2; +pk a +2 2 +commit; +disconnect con1; +connection default; +drop table t1; diff --git a/mysql-test/suite/rocksdb/t/range_locking.test b/mysql-test/suite/rocksdb/t/range_locking.test index 5c21f374bac..e1e8ac92bdc 100644 --- a/mysql-test/suite/rocksdb/t/range_locking.test +++ b/mysql-test/suite/rocksdb/t/range_locking.test @@ -211,3 +211,98 @@ disconnect con1; connection default; drop table t0,t1; +--echo # +--echo # Transaction isolation test +--echo # + +create table t1 (pk int primary key, a int) engine=rocksdb; +insert into t1 values (1,1),(2,2),(3,3); + +connect (con1,localhost,root,,); + +--echo # TRX1: Start, Allocate a snapshot +connection con1; +begin; +select * from t1; + +--echo # TRX2: Make a change that TRX1 will not see +connection default; +update t1 set a=2222 where pk=2; + +--echo # TRX1: Now, make a change that would overwrite TRX2'x change and commit +connection con1; +update t1 set a=a+1 where pk=2; +commit; + +--echo # Examine the result: +--echo # pk=2, a=2223 means UPDATE in TRX1 used "read committed" (InnoDB-like isolation) +--echo # pk=2, a=3 means UPDATE in TRX1 silently overwrote TRX2 +--echo # (and with key tracking, one would get an error on the second UPDATE) +connection default; +select * from t1; + +disconnect con1; +connection default; +drop table t1; + +--echo # +--echo # Same test as above, but check the range scan +--echo # + +create table t1 (pk int primary key, a int) engine=rocksdb; +insert into t1 values (1,1),(2,2),(3,3),(4,4),(5,5),(6,6); + +connect (con1,localhost,root,,); + +--echo # TRX1: Start, Allocate a snapshot +connection con1; +begin; +select * from t1; + +--echo # TRX2: Make a change that TRX1 will not see +connection default; +update t1 set a=2222 where pk between 3 and 5; + +--echo # TRX1: Now, make a change that would overwrite TRX2'x change and commit +connection con1; +update t1 set a=a+1 where pk between 3 and 5; +commit; + +--echo # Examine the result: +--echo # pk={3,4,5} a=2223 means UPDATE in TRX1 used "read committed" (InnoDB-like isolation) +connection default; +select * from t1; + +disconnect con1; +connection default; +drop table t1; + +--echo # +--echo # Same as above, but test SELECT FOR UPDATE. +--echo # +create table t1 (pk int primary key, a int) engine=rocksdb; +insert into t1 values (1,1),(2,2),(3,3),(4,4),(5,5),(6,6); + +connect (con1,localhost,root,,); + +--echo # TRX1: Start, Allocate a snapshot +connection con1; +begin; +select * from t1; + +--echo # TRX2: Make a change that TRX1 will not see +connection default; +update t1 set a=222 where pk=2; +update t1 set a=333 where pk=3; + +--echo # TRX1: Check what select [FOR UPDATE] sees +connection con1; +select * from t1 where pk in (2,3); +select * from t1 where pk=2 for update; +select * from t1 where pk=2; + +commit; + +disconnect con1; +connection default; +drop table t1; diff --git a/storage/rocksdb/ha_rocksdb.cc b/storage/rocksdb/ha_rocksdb.cc index c4f611a663f..f6f11ad4133 100644 --- a/storage/rocksdb/ha_rocksdb.cc +++ b/storage/rocksdb/ha_rocksdb.cc @@ -2101,8 +2101,39 @@ protected: virtual void do_set_savepoint() = 0; virtual void do_rollback_to_savepoint() = 0; + private: + /* + If true, the current statement should not use a snapshot for reading. + Note that in a multi-statement transaction, the snapshot may have been + allocated by another statement. + */ + bool m_stmt_ignores_snapshot = false; + + /* Snapshot-ignore mode will put away m_reads_opts.snapshot here: */ + const rocksdb::Snapshot *m_saved_snapshot; + public: + + void start_ignore_snapshot() { + // note: this may be called several times for the same statement + if (!m_stmt_ignores_snapshot) { + m_saved_snapshot = m_read_opts.snapshot; + m_read_opts.snapshot = nullptr; + m_stmt_ignores_snapshot= true; + } + } + + void end_ignore_snapshot_if_needed() { + if (m_stmt_ignores_snapshot) { + m_stmt_ignores_snapshot = false; + m_read_opts.snapshot = m_saved_snapshot; + m_saved_snapshot = nullptr; + } + } + bool in_snapshot_ignore_mode() const { return m_stmt_ignores_snapshot; } + rocksdb::ReadOptions m_read_opts; + const char *m_mysql_log_file_name; my_off_t m_mysql_log_offset; const char *m_mysql_gtid; @@ -2596,7 +2627,7 @@ public: virtual bool is_tx_started() const = 0; virtual void start_tx() = 0; - virtual void start_stmt() = 0; + virtual void start_stmt(bool is_dml_statement) = 0; void set_initial_savepoint() { /* @@ -2849,7 +2880,7 @@ public: } void acquire_snapshot(bool acquire_now) override { - if (m_read_opts.snapshot == nullptr) { + if (m_read_opts.snapshot == nullptr && !in_snapshot_ignore_mode()) { const auto thd_ss = std::static_pointer_cast<Rdb_explicit_snapshot>( m_thd->get_explicit_snapshot()); if (thd_ss) { @@ -2964,7 +2995,7 @@ public: if (value != nullptr) { value->Reset(); - } + } // psergey-todo: m_read_opts.snapshot below! return m_rocksdb_tx->GetForUpdate(m_read_opts, column_family, key, value, exclusive); } @@ -3028,13 +3059,25 @@ public: /* Start a statement inside a multi-statement transaction. - @todo: are we sure this is called once (and not several times) per - statement start? + @note: If a statement uses N tables, this function will be called N times, + for each TABLE object that is used. For hooking to start of statement that is its own transaction, see ha_rocksdb::external_lock(). */ - void start_stmt() override { + void start_stmt(bool is_dml_statement) override { + + if (rocksdb_use_range_locking && is_dml_statement) { + /* + In Range Locking mode, RocksDB does not do "key tracking". + Use InnoDB-like concurrency mode: make the DML statements always read + the latest data (instead of using transaction's snapshot). + This "downgrades" the transaction isolation to READ-COMMITTED on the + master, but in return the actions can be replayed on the slave. + */ + start_ignore_snapshot(); + } + // Set the snapshot to delayed acquisition (SetSnapshotOnNextOperation) acquire_snapshot(false); } @@ -3270,7 +3313,7 @@ public: set_initial_savepoint(); } - void start_stmt() override {} + void start_stmt(bool is_dml_statement) override {} void rollback_stmt() override { if (m_batch) @@ -3484,8 +3527,10 @@ static int rocksdb_prepare(handlerton *const hton, THD *const thd, DEBUG_SYNC(thd, "rocksdb.prepared"); } - else + else { tx->make_stmt_savepoint_permanent(); + tx->end_ignore_snapshot_if_needed(); + } return HA_EXIT_SUCCESS; } @@ -3660,6 +3705,7 @@ static int rocksdb_commit(handlerton *const hton, THD *const thd, */ tx->set_tx_failed(false); tx->make_stmt_savepoint_permanent(); + tx->end_ignore_snapshot_if_needed(); } if (my_core::thd_tx_isolation(thd) <= ISO_READ_COMMITTED) { @@ -3698,6 +3744,7 @@ static int rocksdb_rollback(handlerton *const hton, THD *const thd, */ tx->rollback_stmt(); + tx->end_ignore_snapshot_if_needed(); tx->set_tx_failed(true); } @@ -4242,13 +4289,19 @@ static bool rocksdb_show_status(handlerton *const hton, THD *const thd, return res; } + +/* + @param is_dml_statement If true, we are is a DML statement +*/ + static inline void rocksdb_register_tx(handlerton *const hton, THD *const thd, - Rdb_transaction *const tx) { + Rdb_transaction *const tx, + bool is_dml_stmt) { DBUG_ASSERT(tx != nullptr); trans_register_ha(thd, FALSE, rocksdb_hton); if (my_core::thd_test_options(thd, OPTION_NOT_AUTOCOMMIT | OPTION_BEGIN)) { - tx->start_stmt(); + tx->start_stmt(is_dml_stmt); trans_register_ha(thd, TRUE, rocksdb_hton); } } @@ -4344,7 +4397,7 @@ static int rocksdb_start_tx_and_assign_read_view( DBUG_ASSERT(!tx->has_snapshot()); tx->set_tx_read_only(true); - rocksdb_register_tx(hton, thd, tx); + rocksdb_register_tx(hton, thd, tx, false); tx->acquire_snapshot(true); if (ss_info) { @@ -4492,7 +4545,7 @@ static int rocksdb_start_tx_with_shared_read_view( DBUG_ASSERT(!tx->has_snapshot()); tx->set_tx_read_only(true); - rocksdb_register_tx(hton, thd, tx); + rocksdb_register_tx(hton, thd, tx, false); tx->acquire_snapshot(true); // case: an explicit snapshot was not assigned to this transaction @@ -11010,7 +11063,7 @@ int ha_rocksdb::external_lock(THD *const thd, int lock_type) { } } tx->m_n_mysql_tables_in_use++; - rocksdb_register_tx(rocksdb_hton, thd, tx); + rocksdb_register_tx(rocksdb_hton, thd, tx, (lock_type == F_WRLCK)); tx->io_perf_start(&m_io_perf); } @@ -11037,7 +11090,7 @@ int ha_rocksdb::start_stmt(THD *const thd, thr_lock_type lock_type) { Rdb_transaction *const tx = get_or_create_tx(thd); read_thd_vars(thd); - rocksdb_register_tx(ht, thd, tx); + rocksdb_register_tx(ht, thd, tx, (lock_type == F_WRLCK)); tx->io_perf_start(&m_io_perf); DBUG_RETURN(HA_EXIT_SUCCESS);