5
Replication events are recorded using messages in the `Google Protocol Buffer
6
<http://code.google.com/p/protobuf/>`_ (GPB) format. GPB messages can contain
7
sub-messages. There is a single main "envelope" message, Transaction, that
8
is passed to plugins that subscribe to the replication stream.
13
**transaction_message_threshold**
15
Controls the size, in bytes, of the Transaction messages. When a Transaction
16
message exceeds this size, a new Transaction message with the same
17
transaction ID will be created to continue the replication events.
18
See :ref:`bulk-operations` below.
23
Controls whether the originating SQL query will be included within each
24
Statement message contained in the enclosing Transaction message. The
25
default global value is FALSE which will not include the query in the
26
messages. It can be controlled per session, as well. For example:
30
drizzle> set @@replicate_query = 1;
32
The stored query should be used as a guide only, and never executed
33
on a slave to perform replication as this will lead to incorrect results.
38
The GPB messages are defined in .proto files in the drizzled/message
39
directory of the Drizzle source code. The primary definition file is
40
transaction.proto. Messages defined in this file are related in the
44
------------------------------------------------------------------
46
| Transaction message |
48
| ----------------------------------------------------------- |
50
| | TransactionContext message | |
52
| ----------------------------------------------------------- |
53
| ----------------------------------------------------------- |
55
| | Statement message 1 | |
57
| ----------------------------------------------------------- |
58
| ----------------------------------------------------------- |
60
| | Statement message 2 | |
62
| ----------------------------------------------------------- |
64
| ----------------------------------------------------------- |
66
| | Statement message N | |
68
| ----------------------------------------------------------- |
69
------------------------------------------------------------------
71
with each Statement message looking like so::
73
------------------------------------------------------------------
77
| ----------------------------------------------------------- |
79
| | Common information | |
81
| | - Type of Statement (INSERT, DELETE, etc) | |
82
| | - Start Timestamp | |
83
| | - End Timestamp | |
84
| | - (OPTIONAL) Actual SQL query string | |
86
| ----------------------------------------------------------- |
87
| ----------------------------------------------------------- |
89
| | Statement subclass message 1 (see below) | |
91
| ----------------------------------------------------------- |
93
| ----------------------------------------------------------- |
95
| | Statement subclass message N (see below) | |
97
| ----------------------------------------------------------- |
98
------------------------------------------------------------------
100
The Transaction Message
101
#######################
103
The main "envelope" message which represents an atomic transaction
104
which changed the state of a server is the Transaction message class.
106
The Transaction message contains two pieces:
108
#. A TransactionContext message containing information about the
109
transaction as a whole, such as the ID of the executing server,
110
the start and end timestamp of the transaction, segmenting
111
metadata and a unique identifier for the transaction.
112
#. A vector of Statement messages representing the distinct SQL
113
statements which modified the state of the server. The Statement
114
message is, itself, a generic envelope message containing a
115
sub-message which describes the specific data modification which
116
occurred on the server (such as, for instance, an INSERT statement).
118
The Statement Message
119
#####################
121
The generic "envelope" message containing information common to each
122
SQL statement executed against a server (such as a start and end timestamp
123
and the type of the SQL statement) as well as a Statement subclass message
124
describing the specific data modification event on the server.
126
Each Statement message contains a type member which indicates how readers
127
of the Statement should construct the inner Statement subclass representing
130
Statements are recorded separately as sometimes individual statements
131
have to be rolled back.
136
How Bulk Operations Work
137
########################
139
Certain operations which change large volumes of data on a server
140
present a specific set of problems for a transaction coordinator or
141
replication service. If all operations must complete atomically on a
142
publishing server before replicas are delivered the complete
145
#. The publishing server could consume a large amount of memory
146
building an in-memory Transaction message containing all the
147
operations contained in the entire transaction.
148
#. A replica, or subscribing server, is wasting time waiting on the
149
eventual completion (commit) of the large transaction on the
150
publishing server. It could be applying pieces of the large
151
transaction in the meantime...
153
In order to prevent the problems inherent in (1) and (2) above, Drizzle's
154
replication system uses a mechanism which provides bulk change
157
A single transaction in the database can possibly be represented with
158
multiple protobuf Transaction messages if the message grows too large.
159
This can happen if you have a bulk transaction, or a single statement
160
affecting a very large number of rows, or just a large transaction with
161
many statements/changes.
163
For the first two examples, it is likely that the Statement sub-message
164
itself will get segmented, causing another Transaction message to be
165
created to hold the rest of the Statement's row changes. In these cases,
166
it is enough to look at the segment information stored in the Statement
167
message (see example below).
169
For the last example, the Statement sub-messages may or may not be
170
segmented, but we could still need to split the individual Statements up into
171
multiple Transaction messages to keep the Transaction message size from
172
growing too large. In this case, the segment information in the Statement
173
submessages is not helpful if the Statement isn't segmented. We need this
174
information in the Transaction message itself.
176
Segmenting a Single SQL Statement
177
*********************************
179
When a regular SQL statement modifies or inserts more rows than a
180
certain threshold, Drizzle's replication services component will begin
181
sending Transaction messages to replicas which contain a chunk
182
(or "segment") of the data which has been changed on the publisher.
184
When data is inserted, updated, or modified in the database, a
185
header containing information about modified tables and fields is
186
matched with one or more data segments which contain the actual
187
values changed in the statement.
189
It's easiest to understand this mechanism by following through a real-world
192
Suppose the following table:
194
.. code-block:: mysql
196
CREATE TABLE test.person
198
id INT NOT NULL AUTO_INCREMENT PRIMARY KEY
199
, first_name VARCHAR(50)
200
, last_name VARCHAR(50)
201
, is_active CHAR(1) NOT NULL DEFAULT 'Y'
204
Also suppose that test.t1 contains 1 million records.
206
Next, suppose a client issues the SQL statement:
208
.. code-block:: mysql
210
UPDATE test.person SET is_active = 'N';
212
It is clear that one million records could be updated by this statement
213
(we say, "could be" since Drizzle does not actually update a record if
214
the UPDATE would not change the existing record...).
216
In order to prevent the publishing server from having to construct an
217
enormous Transaction message, Drizzle's replication services component
218
will do the following:
220
#. Construct a Transaction message with a transaction context containing
221
information about the originating server, the transaction ID, and
222
timestamp information.
223
#. Construct an UpdateHeader message with information about the tables
224
and fields involved in the UPDATE statement. Push this UpdateHeader
225
message onto the Transaction message's statement vector.
226
#. Construct an UpdateData message. Set the *segment_id* member to 1.
227
Set the *end_segment* member to true.
228
#. For every record updated in a storage engine, the ReplicationServices
229
component builds a new UpdateRecord message and appends this message
230
to the aforementioned UpdateData message's record vector.
231
#. After a certain threshold of records is reached, the
232
ReplicationServices component sets the current UpdateData message's
233
*end_segment* member to false, and proceeds to send the Transaction
234
message to replicators.
235
#. The ReplicationServices component then constructs a new Transaction
236
message and constructs a transaction context with the same
237
transaction ID and server information.
238
#. A new UpdateData message is created. The message's *segment_id* is
239
set to N+1 and as new records are updated, new UpdateRecord messages
240
are appended to the UpdateData message's record vector.
241
#. While records are being updated, we repeat steps 5 through 7, with
242
only the final UpdateData message having its *end_segment* member set
245
Segmenting a Transaction
246
************************
248
The Transaction protobuf message also contains *segment_id* member and a
249
*end_segment* member. These values are also set appropriately when a
250
Statement sub-message is segmented, as described above.
252
These values are also set when a Transaction must be segmented along
253
individual Statement boundaries (i.e., the Statement message itself
254
is **not** segmented). In either case, it is enough to check the
255
*end_segment* and *segment_id* values of the Transaction message
256
to determine if this is a multi-message transaction.
261
Both transactions and individual statements may be rolled back.
263
When a transaction is rolled back, one of two things happen depending
264
on whether the transaction is made up of either a single Transaction
265
message, or if it is made up of multiple Transaction messages (e.g, bulk
268
* For a transaction encapsulated entirely within a single Transaction
269
message, the entire message is simply discarded and not sent through
270
the replication stream.
271
* For a transaction which is made up of multiple messages, and at least
272
one message has already been sent through the replication stream, then
273
the Transaction message will contain a Statement message with type =
274
ROLLBACK. This signifies to rollback the entire transaction.
276
A special Statement message type, ROLLBACK_STATEMENT, is used when
277
we have a segmented Statement message (see above) and we need to tell the
278
receiver to undo any changes made for this single statement, but not
279
for the entire transaction. If the receiver cannot handle rolling back
280
a single statement, then a message buffering strategy should be employed
281
to guarantee that a statement was indeed applied successfully before
282
executing on the receiver.
284
.. _replication_streams:
289
The Drizzle kernel handles delivering replication messages to plugins by
290
maintaining a list of replication streams. A stream is represented as a
291
registered *replicator* and *applier* pair.
293
When a replication message is generated within the kernel, the replication
294
services module of the kernel will send this message to each registered
295
*replicator*. The *replicator* will then do something useful with it and
296
send it to each *applier* with which it is associated.
301
A registered *replicator* is a plugin that implements the TransactionReplicator
302
API. Each replicator will be plugged into the kernel to receive the Google
303
Protobuf messages that are generated as the database is changed. Ideally,
304
each registered replicator will transform or modify the messages it receives
305
to implement a specific behavior. For example, filtering by schema name.
307
Each registered replicator should have a unique name. The default replicator,
308
cleverly named **default_replicator**, does no transformation at all on the
309
replication messages.
314
A registered *applier* is a plugin that implements the TransactionApplier
315
API. Appliers are responsible for applying the replication messages that it
316
will receive from a registered replicator. The word "apply" is used loosely
317
here. An applier may do anything with the replication messages that provides
318
useful behavior. For example, an applier may simply write the messages to a
319
file on disk, or it may send the messages over the network to some other
320
service to be processed.
322
At the point of registration with the Drizzle kernel, each applier specifies
323
the name of a registered replicator that it should be attached to in order to
324
make the replication stream pair.