WL#4162: Add start/end of trx/stm to binlog
Affects: Server-5.7 — Status: In-Design — Priority: Medium
Executive Summary ================= This worklog provides information at each event logged to binary log about transactions boundaries by setting markers in the logged event. These markers use three single bits of the Log_event Common-Header flags. By using single bits of the Log_event Common-Header flags this feature will need no additional space in binary log to store these markers. With such markers, by reading a single Log_event header from an event stream, it will be possible to state if the event: a) is from a server with the feature enabled; b) has no transaction boundary information; c) is a self contained event; d) is starting a transaction; e) is inside a transaction; f) is ending a transaction; g) is in the middle of a transaction; h) should not be considered in the transaction boundary parsing. Before this worklog it was needed to parse a set of events in an event stream to get information about transactions boundaries. The old method will still be used when dealing with events without boundary markers. After this worklog, this feature will be always enabled. There will be no option to disable tagging of transaction boundary markers in MySQL server. References ========== Bug#17943188 BUG#26395, WL#3735
Functional requirements ======================= F1) Markers should be written to the binary log. F2) Markers should be visible as comments in mysqlbinlog output. F3) BINLOG statements inner events are left untouched in respect to transaction boundary markers. Non-Functional requirements =========================== NF1) The feature must be supported regardless the use of a specific binary log format. NF2) The feature must be supported regardless the use of GTIDs.
Concepts and Definitions ======================== C1) Transaction Stream/Transaction Event Group/Commit Group: a sequence of events related to a single transaction. D1) TBM: Transaction Boundaries Markers. The bits in the binary log event common-header in with a transaction boundary type will be written. D2) Transaction Boundary Type: the type of transaction boundary of a single event in a server with the transaction boundaries feature enabled. Of the eight possible types (three bits), this worklog is using six, letting the remaining two types available for future use. The transaction boundary types defined in this worklog are: D2.0) Not Defined: The event was logged by a server without the feature. D2.1) Ignore Boundary: The event can appear in the middle of a transaction and should be ignored by the transaction boundary parser. This is the case of fake Rotate events sent by the dump_thread when a slave connects asking data. Examples: Format_description, Rotate. D2.2) Self Contained: Any event that is self contained (doesn't depend on previous or following events). Examples: D2.1a) Any Query event with an auto-commit statement (DDL for example) if GTIDs are disabled and without Intvar, Rand or User_var related events; D2.1b) Start, Stop, Rotate, Previous_gtids, etc. D2.3) Start Transaction: Any event starting a Transaction Stream. D2.4) Inside Transaction: Any event that is part of a Transaction Stream between the event that starts and the event that ends the stream. D2.5) End Transaction: Any event ending a Transaction Stream. High-Level Specification ======================== Boundary Type ------------- Represents a transaction boundary type. Every Log_event object have a boundary type. The boundary type for an event will be determined by the event type, but it can be changed before writing to binary log depending on: a) Use of GTIDs; b) Dependency on Intvar, Rand or User_var events; A slave I/O thread reading events from the master can use the event's boundary type to verify if the event is the middle of a Transaction Stream. This will be important to WL#7349 and WL#7355. Transaction Boundary Parser --------------------------- Responsible for keeping track of transactions boundaries in an event stream and also for assuring that the event stream is valid. A server with this feature enabled will expect that a valid stream of events follows the rules defined below: BOUNDARY_TYPE_NOT_DEFINED ( (BOUNDARY_TYPE_SELF_CONTAINED|BOUNDARY_TYPE_IGNORE_BOUNDARY)* BOUNDARY_TYPE_START_TRANSACTION (BOUNDARY_TYPE_INSIDE_TRANSACTION|BOUNDARY_TYPE_IGNORE_BOUNDARY)* BOUNDARY_TYPE_END_TRANSACTION )* Where (<boundary type>)* means 0 or more repetitions of the types. The Transaction Boundary Parser have implemented the rules to support the verification of a event stream either by a slave I/O thread reading events from the master or by the mysqlbinlog client program reading events from a binlog file. Transaction Boundary Parser Transition Table ==================================================================== |From boundary type ->| 0 | 2 | 3 | 4 | 5 | | | Not |Self Con| Start | Inside | End | |To boundary type |Defined | tained | Trans | Trans | Trans | |=====================|========|========|========|========|========| |0) Not Defined | This is the parser starting type |*1 |---------------------|--------|--------|--------|--------|--------| |1) Ignore Boundary | The parser is never set to this type |*2 |---------------------|--------|--------|--------|--------|--------| |2) Self Contained | Ok | Ok | Err | Err | Ok | |---------------------|--------|--------|--------|--------|--------| |3) Start Transaction | Ok | Ok | Err | Err | Ok | |---------------------|--------|--------|--------|--------|--------| |4) Inside Transaction| Err | Err | Ok | Ok | Err | |---------------------|--------|--------|--------|--------|--------| |5) End Transaction | Err | Err | Ok | Ok | Err | ==================================================================== *1 - When feeding the transaction boundary parser with an event with BOUNDARY_TYPE_NOT_DEFINED in its markers, the transaction boundary parser will define (but will not mark) the event boundary type based on same assumptions the master do for marking. So, we can have a server without this feature replicating to a slave with the feature and, instead of only reading the markers (the easy way), the slave will "guess" the boundary type of the events to take benefit of the feature. *2 - When feeding the transaction boundary parser with an event with BOUNDARY_TYPE_IGNORE_BOUNDARY in its markers, the transaction boundary parser will keep its current type. A transaction boundary parser must be "reset()" every time the server move from a current event stream to another one (example: when the slave server connects to a new master). This reset procedure will move the boundary type of the parser to BOUNDARY_TYPE_NOT_DEFINED. When reading events from a master server (or from a binlog file), the Master_info boundary parser (or the mysqlbinlog client program's one) will be fed with the buffers of the event (this contains the event header and, off course, the transaction boundary markers). If the stream is fed with an event that result in an incorrect transition in the state machine (examples: Self -> End Transaction or End Transaction -> Inside Transaction), a warning/error message I2/I3 described in the Interface Specification will be logged. What this message means to the user: if a slave connects to a master and asks events pointing an binlog position in the middle of a transaction, the error will be thrown (probably stating that it was "Unable to change boundary parser from Not Defined to Inside Transaction)". How can the user fix things so that he does not get the message again: In the example above, the user can check where the transaction is starting (using mysqlbinlog client program) and try to position the slave connection correctly using CHANGE MASTER TO. Interface Specification ======================= I1) mysqlbinlog client program dump will show the boundary type for the events. I2) A new warning message can be logged by the slave server if the I/O thread fails to feed an event into the transaction boundary parser. This could happen if the event boundary type breaks the rules of the boundary parser. It will also make an assert to fail in debug builds. I3) A new warning message can be logged by mysqlbinlog client program if a break in the transaction boundary parser rules happens when an event from an event stream is fed.
Low-Level Design Specification ============================== Transaction boundary types for log event types ============================================== Log_event type |State | Set by ---------------------------------+------+------------------------------ Log_event |SELF | Log_event constructor Query_log_event |SELF | Log_event constructor *1 BEGIN |START | Query_log_event constructor COMMIT/ROLLBACK |END | Query_log_event constructor not immediate logging |INSIDE| Query_log_event constructor BEGIN |INSIDE| binlog write_event *2 immediate logging |END | binlog write_event *2 Execute_load_query_log_event | | Query_log_event constructor*3 Load_log_event |SELF | Log_event constructor *1 Create_file_log_event |SELF | Log_event constructor *1 Start_log_event_v3 |SELF | Log_event constructor *1 Format_description_log_event |IGNORE| Log_event constructor Intvar_log_event |INSIDE| Intvar constructor Intvar_log_event |START | binlog write_event *4 Rand_log_event |INSIDE| Rand_log_event constructor Rand_log_event |START | binlog write_event *4 User_var_log_event |INSIDE| User_var constructor User_var_log_event |START | binlog write_event *4 Xid_log_event |END | Xid_log_event constructor Stop_log_event |SELF | Log_event constructor *1 Rotate_log_event |SELF | Log_event constructor *1 Rotate_log_event |IGNORE| Log_event constructor *5 Append_block_log_event |INSIDE| Append_block... constructor Begin_load_query_log_event |INSIDE| Append_block... constructor Delete_file_log_event |SELF | Log_event constructor *1 Execute_load_log_event |SELF | Log_event constructor *1 Unknown_log_event |SELF | Log_event constructor *1 Table_map_log_event |INSIDE| Table_map constructor Rows_log_event |INSIDE| Rows_log_event constructor Write_rows_log_event |INSIDE| Rows_log_event constructor Update_rows_log_event |INSIDE| Rows_log_event constructor Delete_rows_log_event |INSIDE| Rows_log_event constructor Incident_log_event |SELF | Log_event constructor *1 Ignorable_log_event |IGNORE| Ignorable_log... constructor Rows_query_log_event |INSIDE| Rows_query_log... constructor Heartbeat_log_event |SELF | Log_event constructor *1 Gtid_log_event |START | Gtid_log_event constructor Previous_gtid_log_event |SELF | Log_event constructor *1 ======================================================================= *1 - No special code was added to these events. The boundary type is "inherited" from Log_event. *2 - See binlog.cc modification M1. *3 - The state will be set by Query_log_event constructor. *4 - See binlog.cc modification M2. *5 - See rpl_binlog_sender.cc modification M1. Note: not sure yet about boundary type for Incident_log_event. Modifications in source files ============================= rpl_trx_boundary_parser.h/rpl_trx_boundary_parser.cc (new files) ---------------------------------------------------------------- M1) Defined the boundary markers macros for dealing with event flags bitwise operations, the possible boundary types and their names. M2) Defined the Transaction_boundary_parser class. log_event.h ----------- M1) Added an include to rpl_trx_boundary_parser.h M2) Log_event class definition: M2a) Added a public boundary type named boundary_type to the Log_event class definition; M2b) Created a new protected method called update_boundary_type_from_flags() to update the boundary type of the event with the values of the Log_event flags. M3) Query_log_event class definition: M3a) Created a new method is_begin_stmt(); M3b) Created a new method is_commit_stmt(); M3c) Created a new method is_rollback_stmt(); M3d) Changed starts_group() method to use is_begin_stmt(); M3e) Changed ends_group() method to use is_commit_stmt() and is_rollback_stmt(). M4) Intvar_log_event constructor: Added code to mark the event as BOUNDARY_TYPE_INSIDE_TRANSACTION. M5) Rand_log_event constructor: Added code to mark the event as BOUNDARY_TYPE_INSIDE_TRANSACTION. M6) Xid_log_event constructor: Added code to mark the event as BOUNDARY_TYPE_END_TRANSACTION. M7) User_var_log_event constructor: Added code to mark the event as BOUNDARY_TYPE_INSIDE_TRANSACTION. M8) Rows_query_log_event constructor: Added code to mark the event as BOUNDARY_TYPE_INSIDE_TRANSACTION. log_event.cc ------------ M1) Log_event constructors: Added code to mark the event as BOUNDARY_TYPE_SELF_CONTAINED. M2) Log_event::write_header: Added code to apply the boundary type into the flags of the event before writing it. M3) Log_event::read_log_event: Added code to set event's boundary type according to event's flags. Obs.: The *_event constructors that have code to set their boundary types are those not used to reconstruct a previously logged event. When the event is being reconstructed (like using read_log_event), the boundary type is set by this modification. Changed a debug message to show also event's boundary type. M4) Log_event::print_header: Prints the boundary type of the event. M5) Query_log_event constructor: Before the verification of 'ignore_cmd_internals', we adjust the event's boundary type to BOUNDARY_TYPE_INSIDE_TRANSACTION if it is a BEGIN statement or to BOUNDARY_TYPE_END_TRANSACTION if it is a COMMIT or ROLLBACK statement. Later, if the Query event is not using immediate logging, we set the event boundary type to BOUNDARY_TYPE_INSIDE_TRANSACTION. M6) Append_block_log_event constructor: Added code to mark the event as BOUNDARY_TYPE_INSIDE_TRANSACTION. M7) Rows_log_event constructor: Added code to mark the event as BOUNDARY_TYPE_INSIDE_TRANSACTION. M8) Table_map_log_event constructor: Added code to mark the event as BOUNDARY_TYPE_INSIDE_TRANSACTION. M9) Gtid_log_event constructor: Added code to mark the event as BOUNDARY_TYPE_START_TRANSACTION. binlog.cc --------- M1) binlog_cache_data::write_event: Every Gtid_log_event is marked as BOUNDARY_TYPE_START_TRANSACTION by Gtid_log_event's constructor. If a Gtid_log_event have to be logged before the event to be written, it is needed to adjust the boundary type of the event to be written as follows: M1a) If ev->boundary_type is BOUNDARY_TYPE_START_TRANSACTION, we have to mark it as BOUNDARY_TYPE_INSIDE_TRANSACTION; M1b) If ev->boundary_type is BOUNDARY_TYPE_SELF_CONTAINED, we have to mark it as BOUNDARY_TYPE_END_TRANSACTION. M2) MYSQL_BIN_LOG::write_event: Here it is possible to insert Intvar_log_event, Rand_log_event or User_var_log_event if needed before writing the event asked to be written. As all Intvar_log_event, Rand_log_event and User_var_log_event are marked as BOUNDARY_TYPE_INSIDE_TRANSACTION by their constructors, if the event to be written is marked as BOUNDARY_TYPE_SELF_CONTAINED (it is a DDL, for example), we change the first inserted event markers to BOUNDARY_TYPE_START_TRANSACTION and the event to be written is marked as BOUNDARY_TYPE_END_TRANSACTION. rpl_binlog_sender.cc -------------------- M1) Binlog_sender::fake_rotate_event: This method will send a rotate log_event without instantiating an Rotate_log_event object. The flags of the fake rotate event to be sent is marked as BOUNDARY_TYPE_IGNORE_BOUNDARY. M2) Binlog_sender::send_heartbeat_event: This method will send a heartbeat log_event without instantiating an Heartbeat_log_event object. The flags of the heartbeat event to be sent is marked as BOUNDARY_TYPE_SELF_CONTAINED. rpl_mi.h -------- M1) Added an include to rpl_trx_boundary_parser.h M2) Master_info class definition: Added an Transaction_boundary_parser object named event_parser to the Master_info class definition. rpl_slave.cc ------------ M1) Added an include to rpl_trx_boundary_parser.h M2) handle_slave_io: Added event boundary type to a debug "info" message stating: "IO thread received event..." Added a call to mi->tbm_state_machine.reset() right before entering the loop reading the events. M3) queue_event: Added code to feed the Master_info event_parser with the the buffer of the event to being queued. If the mi->event_parser returns that it was an invalid transition, throw an warning message stating: "Unable to change boundary parser from %s to %s." and the boundary type names of the previous boundary type in the state machine and the last boundary type which the state machine was fed. After throwing the warning message, as the event_parser should now be in the error state, we will reset() the event parser, moving it to the "Not Defined" type. For the debug builds, we will add an assert to verify if the even_event parser has changed to a type other than error. For non debug builds, the I/O thread will throw a warning message on each event queued until it reaches a new transaction boundary parser consistent state (a self contained event or a start transaction). mysqlbinlog.cc -------------- M1) Added an include to rpl_trx_boundary_parser.h M2) Added an Transaction_boundary_parser object named event_parser. M3) dump_local_log_entries: M3a) Added a call to event_parser.reset() to initialize the verification of the event stream to be read. M3b) Added code to feed the event_parser with the buffer of the event read. If the event_parser returns that it was an invalid transition, do two things: M3b1) show a warning message stating: "Unable to change boundary parser from %s to %s." and the boundary type names of the previous boundary type and the boundary type of the event read; M3b2) Issue an event_parser.reset() to remove the transaction boundary parser from the error state. M4) Added an include to rpl_trx_boundary_parser.cc at the end of the file.
Copyright (c) 2000, 2016, Oracle Corporation and/or its affiliates. All rights reserved.