MySQL Cluster Manager 1.3.5 is now available to download from My Oracle Support.
Details are available in the the MCM 1.3.5 Release Notes. Note that this version of MCM now supports MySQL Cluster 7.4 (as well as earlier versions or MySQL Cluster).
MySQL Cluster Manager 1.3.5 is now available to download from My Oracle Support.
Details are available in the the MCM 1.3.5 Release Notes. Note that this version of MCM now supports MySQL Cluster 7.4 (as well as earlier versions or MySQL Cluster).
The binary and source versions of MySQL Cluster 7.4.6 have now been made available at http://www.mysql.com/downloads/cluster/.
MySQL Cluster NDB 7.4.6 is a new maintenance release of MySQL Cluster, based on MySQL Server 5.6 and including features from version 7.4 of the NDB storage engine, as well as fixing a number of recently discovered bugs in previous MySQL Cluster releases.
This release also incorporates all bugfixes and changes made in previous MySQL Cluster releases, as well as all bugfixes and feature changes which were added in mainline MySQL 5.6 through MySQL 5.6.24.
On Thursday I’ll be hosting a webinar introducing MySQL Cluster and then looking into what’s new in the latest version (MySQL Cluster 7.4) in some more detail. As always the webinar is free but please register here. I’ll be answering on-line Q&A during the presentation.
Even if you can’t join the live webinar, it’s worth registering as you’ll be emailed a link to the replay as soon as it’s available.
MySQL Cluster powers the subscriber databases of major communication services providers as well as next generation web, cloud, social and mobile applications. It is designed to deliver:
Understand some of the highlights of MySQL Cluster 7.4:
Last week I hosted a webinar explaining how you can get the best from the NoSQL world while still getting all of the benefits of a proven RDBMS. The webinar replay is now available to view here.
There’s a lot of excitement about NoSQL data stores, with the promise of simple access patterns, flexible schemas, scalability, and high availability. The downside comes in the form of losing ACID transactions, consistency, flexible queries, and data integrity checks. What if you could have the best of both worlds? Join this webinar to learn how MySQL Cluster provides simultaneous SQL and native NoSQL access to your data—whether it’s in a simple key-value API (memcached) or REST, JavaScript, Java, or C++. You will hear how the MySQL Cluster architecture delivers in-memory real-time performance; 99.999 percent availability; online maintenance; and linear, horizontal scalability through transparent autosharding.
On Thursday I’ll be hosting a webinar explaining how you can get the best from the NoSQL world while still getting all of the benefits of a proven RDBMS. As always the webinar is free but please register here.
Even if you can’t join the live webinar, it’s worth registering as you’ll be emailed a link to the replay as soon as it’s available.
There’s a lot of excitement about NoSQL data stores, with the promise of simple access patterns, flexible schemas, scalability, and high availability. The downside comes in the form of losing ACID transactions, consistency, flexible queries, and data integrity checks. What if you could have the best of both worlds? Join this webinar to learn how MySQL Cluster provides simultaneous SQL and native NoSQL access to your data—whether it’s in a simple key-value API (memcached) or REST, JavaScript, Java, or C++. You will hear how the MySQL Cluster architecture delivers in-memory real-time performance; 99.999 percent availability; online maintenance; and linear, horizontal scalability through transparent autosharding.
The binary and source versions of MySQL Cluster 7.4.5 have now been made available at http://www.mysql.com/downloads/cluster/.
MySQL Cluster NDB 7.4.5 is a new maintenance release of MySQL Cluster, based on MySQL Server 5.6 and including features from version 7.4 of the NDB storage engine, as well as fixing a number of recently discovered bugs in previous MySQL Cluster releases.
This release also incorporates all bugfixes and changes made in previous MySQL Cluster releases, as well as all bugfixes and feature changes which were added in mainline MySQL 5.6 through MySQL 5.6.23.
It was found during testing that problems could arise when the
node registered as the arbitrator disconnected or failed during
the arbitration process.
In this situation, the node requesting arbitration could never
receive a positive acknowledgement from the registered
arbitrator; this node also lacked a stable set of members and
could not initiate selection of a new arbitrator.
Now in such cases, when the arbitrator fails or loses contact
during arbitration, the requesting node immediately fails rather
than waiting to time out.
(Bug #20538179)
DROP DATABASE
failed to remove
the database when the database directory contained a
.ndb
file which had no corresponding table
in NDB
. Now, when executing
DROP DATABASE
, NDB
performs an check specifically for leftover
.ndb
files, and deletes any that it finds.
(Bug #20480035)
References: See also Bug #44529.
The maximum failure time calculation used to ensure that normal
node failure handling mechanisms are given time to handle
survivable cluster failures (before global checkpoint watchdog
mechanisms start to kill nodes due to GCP delays) was
excessively conservative, and neglected to consider that there
can be at most number_of_data_nodes
/
NoOfReplicas
node
failures before the cluster can no longer survive. Now the value
of NoOfReplicas
is properly taken into
account when performing this calculation.
(Bug #20069617, Bug #20069624)
References: See also Bug #19858151, Bug #20128256, Bug #20135976.
During a node restart, if there was no global checkpoint
completed between the START_LCP_REQ
for a
local checkpoint and its LCP_COMPLETE_REP
it
was possible for a comparison of the LCP ID sent in the
LCP_COMPLETE_REP
signal with the internal
value SYSFILE->latestLCP_ID
to fail.
(Bug #76113, Bug #20631645)
When sending LCP_FRAG_ORD
signals as part of
master takeover, it is possible that the master not is not
synchronized with complete accuracy in real time, so that some
signals must be dropped. During this time, the master can send a
LCP_FRAG_ORD
signal with its
lastFragmentFlag
set even after the local
checkpoint has been completed. This enhancement causes this flag
to persist until the statrt of the next local checkpoint, which
causes these signals to be dropped as well.
This change affects ndbd only; the issue
described did not occur with ndbmtd.
(Bug #75964, Bug #20567730)
When reading and copying transporter short signal data, it was
possible for the data to be copied back to the same signal with
overlapping memory.
(Bug #75930, Bug #20553247)
NDB node takeover code made the assumption that there would be
only one takeover record when starting a takeover, based on the
further assumption that the master node could never perform
copying of fragments. However, this is not the case in a system
restart, where a master node can have stale data and so need to
perform such copying to bring itself up to date.
(Bug #75919, Bug #20546899)
Cluster API:
A scan operation, whether it is a single table scan or a query
scan used by a pushed join, stores the result set in a buffer.
This maximum size of this buffer is calculated and preallocated
before the scan operation is started. This buffer may consume a
considerable amount of memory; in some cases we observed a 2 GB
buffer footprint in tests that executed 100 parallel scans with
2 single-threaded (ndbd) data nodes. This
memory consumption was found to scale linearly with additional
fragments.
A number of root causes, listed here, were discovered that led
to this problem:
Result rows were unpacked to full
NdbRecord
format before
they were stored in the buffer. If only some but not all
columns of a table were selected, the buffer contained empty
space (essentially wasted).
Due to the buffer format being unpacked,
VARCHAR
and
VARBINARY
columns always had
to be allocated for the maximum size defined for such
columns.
BatchByteSize
and
MaxScanBatchSize
values were not taken into consideration as a limiting
factor when calculating the maximum buffer size.
These issues became more evident in NDB 7.2 and later MySQL
Cluster release series. This was due to the fact buffer size is
scaled by BatchSize
, and
that the default value for this parameter was increased fourfold
(from 64 to 256) beginning with MySQL Cluster NDB 7.2.1.
This fix causes result rows to be buffered using the packed
format instead of the unpacked format; a buffered scan result
row is now not unpacked until it becomes the current row. In
addition, BatchByteSize
and
MaxScanBatchSize
are now used as limiting
factors when calculating the required buffer size.
Also as part of this fix, refactoring has been done to separate
handling of buffered (packed) from handling of unbuffered result
sets, and to remove code that had been unused since NDB 7.0 or
earlier. The NdbRecord
class declaration has
also been cleaned up by removing a number of unused or redundant
member variables.
(Bug #73781, Bug #75599, Bug #19631350, Bug #20408733)
MySQL Cluster Manager 1.3.4 is now available to download from My Oracle Support and from the Oracle Software Delivery Cloud.
Details are available in the the MCM 1.3.4 Release Notes. Note that this version of MCM now supports MySQL Cluster 7.4 (as well as earlier versions or MySQL Cluster).
The General Availability of MySQL Cluster 7.4 has just been announced by Oracle.
The MySQL team at Oracle are excited to announce the General Availability of MySQL Cluster 7.4, ready for production workloads.
MySQL Cluster 7.4.4 can be downloaded from mysql.com and the release notes viewed in the MySQL docs.
Figure 1 provides a summary of the enhancements delivered in this release:
The sections that follow delve into some more details.
Being a scaled-out, in-memory, real-time database, MySQL Cluster performance has always been great but we continue to work on making it faster each release. In particular, we want to keep pace with the trend of having more and more cores rather than faster ones. 7.4 continues along the path of better exploiting multiple cores – as can be seen from these benchmark results:
Just make sure that you’re using the multi-threaded data node (ndbmtd rather than ndbd) and have configured how many threads it should use.
The tests were performed using a 48 core/96 thread machine (also demonstrating how well MySQL Cluster can now scale up with large numbers of cores).
So, the sysbench benchmark results show how much faster MySQL Cluster 7.4 is compared to the previous 2 releases (a nice apples-to-apples test but on a very small cluster without any scale-out) but we now have some even more interesting benchmark results showing just how well MySQL Cluster can scale out to deliver incredible results…
Using the flexAsynch benchmark tool and a pool of 32 data nodes (with each data node running on a dedicated 56 thread Intel E5-2697 v3 (Haswell) machine) the cluster was able to process 200,000,000 reads per second. These results are shown in Figure 4, where the number of data nodes are scaled on the x-axis – as you’ll observe, performance scales virtually linearly with the number of data nodes. Note that MySQL Cluster supports up to 48 data nodes and so there’s still plenty of scope to scale out even further.
Using a smaller cluster (just 16 data nodes running on the same 56 thread machines), the DBT2 Benchmark has been used to assess how well SQL performance scales as more data nodes are added. SQL performance was measured at 2,500,000 SQL read opeations per minute. This equates to around 5 Million Transactions Per Minutes or 2.2 Million NewOnly TPM.
As can be in Figure 5 the scaling of SQL reads is again almost linear.
MySQL Cluster allows bi-directional replication between two (or more) clusters. Replication within each cluster is synchronous but between clusters it is asynchronous which means the following scenario is possible:
Site A | Replication | Site B |
---|---|---|
x == 10 | x == 10 | |
x = 11 | x = 20 | |
— x=11 –> | x == 11 | |
x==20 | <– x=20 — |
In this example a value (column for a row in a table) is set to 11 on site A and the change is queued for replication to site B. In the mean time, an application sets the value to 20 on site B and that change is queued for replication to site A. Once both sites have received and applied the replicated change from the other cluster site A contains the value 20 while site B contains 11 – in other words the databases are now inconsistent.
There are two phases to establishing consistency between both clusters after an inconsistency has been introduced:
The following animation illustrates how MySQL Cluster 7.2 detects that an inconsistency has been introduced by the asynchronous, active-active replication:
While we typically consider the 2 clusters in an active-active replication configuration to be peers, in this case we designate one to be the primary and the other the secondary. Reads and writes can still be sent to either cluster but it is the responsibility of the primary to identify that a conflict has arisen and then remove the inconsistency.
A logical clock is used to identify (in relative terms) when a change is made on the primary – for those who know something of the MySQL Cluster internals, we use the index of the Global Checkpoint that the update is contained in. For all tables that have this feature turned on, an extra, hidden column is automatically added on the primary – this represents the value of the logical clock when the change was made.
Once the change has been applied on the primary, there is a “window of conflict” for the effected row(s) during which if a different change is made to the same row(s) on the secondary then there will be an inconsistency. Once the slave on the secondary has applied the change from the primary, it will send a replication event back to the slave on the primary, containing the primary’s clock value associated with the changes that have just been applied on the secondary. (Remember that the clock is actually the Global Checkpoint Index and so this feature is sometimes referred to as Reflected GCI). Once the slave on the primary has received this event, it knows that all changes tagged with a clock value no later than the reflected GCI are now safe – the window of conflict has closed.
If an application modifies this same row on the secondary before the replication event from the primary was applied then it will send an associated replication event to the slave on the primary before it reflects the new GCI. The slave on the primary will process this replication event and compare the clock value recorded with the effected rows with the latest reflected GCI; as the clock value for the conflicting row is higher the primary recognises that a conflict has occured and will launch the algorithm to resolve the inconsistency.
After a conflict has been detected, you have the option of having the database simply report the conflict to the application or have it roll back just the conflicting row or the entire transaction and all subsequent transactions that were dependent on it.
So – what’s new in 7.4?
This section assumes that replication has already been set up between two clusters as shown in Figure 8. For more details on how to set up that configuration then refer to the blog: Enhanced conflict resolution with MySQL Cluster active-active replication.
To keep things simple, just two hosts are used; 192.168.56.101 contains all nodes for the primary cluster and 192.168.56.102 will contain all nodes for the secondary. A single MySQL Server in each cluster acts as both the master and the slave.
The first step is to identify the tables that need conflict detection enabling. Each of those tables then has to have an entry in the mysql.ndb_replication table where they’re tagged as using the new NDB$EPOCH2_TRANS() function – you could also choose to use NDB$EPOCH(), in which case only the changes to conflicting rows will be rolled-back rather than the full transactions. A few things to note:
PRIMARY> CREATE TABLE mysql.ndb_replication ( -> db VARBINARY(63), -> table_name VARBINARY(63), -> server_id INT UNSIGNED, -> binlog_type INT UNSIGNED, -> conflict_fn VARBINARY(128), -> PRIMARY KEY USING HASH (db, table_name, server_id) -> ) ENGINE=NDB -> PARTITION BY KEY(db,table_name); PRIMARY> REPLACE INTO mysql.ndb_replication VALUES ('clusterdb', 'simple1', 0, 0, 'NDB$EPOCH2_TRANS()'); PRIMARY> REPLACE INTO mysql.ndb_replication VALUES ('clusterdb', 'simple2', 0, 0,'NDB$EPOCH2_TRANS()'); PRIMARY> REPLACE INTO mysql.ndb_replication VALUES ('clusterdb', 'simple3', 0, 0,'NDB$EPOCH2_TRANS()'); SECONDARY> SELECT * FROM mysql.ndb_replication; +-----------+------------+-----------+-------------+--------------------+ | db | table_name | server_id | binlog_type | conflict_fn | +-----------+------------+-----------+-------------+--------------------+ | clusterdb | simple2 | 0 | 0 | NDB$EPOCH2_TRANS() | | clusterdb | simple1 | 0 | 0 | NDB$EPOCH2_TRANS() | | clusterdb | simple3 | 0 | 0 | NDB$EPOCH2_TRANS() | +-----------+------------+-----------+-------------+--------------------+
Refer to the MySQL Cluster documentation for details on the contents of the mysql.ndb_replication table.
The next step is to define which MySQL Cluster instance is to act as the PRIMARY and which as the SECONDARY (note that the slave SQL thread must be halted to make this change):
PRIMARY> STOP SLAVE SQL_THREAD; SECONDARY> STOP SLAVE SQL_THREAD; PRIMARY> SET GLOBAL ndb_slave_conflict_role='PRIMARY'; SECONDARY> SET GLOBAL ndb_slave_conflict_role='SECONDARY'; SECONDARY> START SLAVE SQL_THREAD; PRIMARY> START SLAVE SQL_THREAD;
For each of these tables you should also create an exceptions table which will record any conflicts that have resulted in changes being rolled back; the format of these tables is rigidly defined and so take care to copy the types exactly; again this only needs doing on the primary:
PRIMARY> CREATE DATABASE clusterdb;USE clusterdb; PRIMARY> CREATE TABLE simple1$EX ( NDB$server_id INT UNSIGNED, NDB$master_server_id INT UNSIGNED, NDB$master_epoch BIGINT UNSIGNED, NDB$count INT UNSIGNED, NDB$OP_TYPE ENUM('WRITE_ROW','UPDATE_ROW', 'DELETE_ROW', 'REFRESH_ROW', 'READ_ROW') NOT NULL, NDB$CFT_CAUSE ENUM('ROW_DOES_NOT_EXIST', 'ROW_ALREADY_EXISTS', 'DATA_IN_CONFLICT', 'TRANS_IN_CONFLICT') NOT NULL, NDB$ORIG_TRANSID BIGINT UNSIGNED NOT NULL, id INT NOT NULL, value$OLD INT, value$NEW INT, PRIMARY KEY(NDB$server_id, NDB$master_server_id, NDB$master_epoch, NDB$count)) ENGINE=NDB; PRIMARY> CREATE TABLE simple2$EX ( NDB$server_id INT UNSIGNED, NDB$master_server_id INT UNSIGNED, NDB$master_epoch BIGINT UNSIGNED, NDB$count INT UNSIGNED, NDB$OP_TYPE ENUM('WRITE_ROW','UPDATE_ROW', 'DELETE_ROW', 'REFRESH_ROW', 'READ_ROW') NOT NULL, NDB$CFT_CAUSE ENUM('ROW_DOES_NOT_EXIST', 'ROW_ALREADY_EXISTS', 'DATA_IN_CONFLICT', 'TRANS_IN_CONFLICT') NOT NULL, NDB$ORIG_TRANSID BIGINT UNSIGNED NOT NULL, id INT NOT NULL, value$OLD INT, value$NEW INT, PRIMARY KEY(NDB$server_id, NDB$master_server_id, NDB$master_epoch, NDB$count)) ENGINE=NDB; PRIMARY> CREATE TABLE simple3$EX ( NDB$server_id INT UNSIGNED, NDB$master_server_id INT UNSIGNED, NDB$master_epoch BIGINT UNSIGNED, NDB$count INT UNSIGNED, NDB$OP_TYPE ENUM('WRITE_ROW','UPDATE_ROW', 'DELETE_ROW', 'REFRESH_ROW', 'READ_ROW') NOT NULL, NDB$CFT_CAUSE ENUM('ROW_DOES_NOT_EXIST', 'ROW_ALREADY_EXISTS', 'DATA_IN_CONFLICT', 'TRANS_IN_CONFLICT') NOT NULL, NDB$ORIG_TRANSID BIGINT UNSIGNED NOT NULL, id INT NOT NULL, value$OLD INT, value$NEW INT, PRIMARY KEY(NDB$server_id, NDB$master_server_id, NDB$master_epoch, NDB$count)) ENGINE=NDB;
Refer to the MySQL Cluster documentation for details on the format of the exception tables.
Finally, the application tables themselves can be created (this only needs doing on the primary as they’ll be replicated to the secondary):
PRIMARY> CREATE TABLE simple1 (id INT NOT NULL PRIMARY KEY, value INT) ENGINE=ndb; PRIMARY> CREATE TABLE simple2 (id INT NOT NULL PRIMARY KEY, value INT) ENGINE=ndb; PRIMARY> CREATE TABLE simple3 (id INT NOT NULL PRIMARY KEY, value INT) ENGINE=ndb;
Everything is now set up and the new configuration can be tested to ensure that conflicts are detected and the correct updates are rolled back.
The first step is to add some data to our new tables (note that at this point replication is running and so they only need to be created on the primary) and confirm that everything is replicated to the secondary:
PRIMARY> REPLACE INTO simple1 VALUES (1,10); PRIMARY> REPLACE INTO simple2 VALUES (1,10); PRIMARY> REPLACE INTO simple3 VALUES (1,10); SECONDARY> SELECT * FROM simple1; +----+-------+ | id | value | +----+-------+ | 1 | 10 | +----+-------+ SECONDARY> SELECT * FROM simple2; +----+-------+ | id | value | +----+-------+ | 1 | 10 | +----+-------+ SECONDARY> SELECT * FROM simple3; +----+-------+ | id | value | +----+-------+ | 1 | 10 | +----+-------+
It is important that the NDB$EPOCH2_TRANS() function rolls back any transactions on the secondary that involve a conflict (as well as subsequent, dependent transactions that modify the same rows); to intentionally trigger this, the simplest approach is to stop the slave IO thread on the secondary Cluster in order to increase the size of the window of conflict (which is otherwise very short). Once the slave IO thread has been stopped a change is made to table simple1 on the primary and then the secondary makes a (conflicting) change to the same row as well as making a change to table simple2 in the same transaction. A second transaction on the primary will change a row in simple3 – as it doesn’t touch any rows that have been involved in a conflict then that change should stand.
SECONDARY> STOP SLAVE IO_THREAD; PRIMARY> UPDATE simple1 SET value=13 WHERE id=1; SECONDARY> BEGIN; # conflicting transaction SECONDARY> UPDATE simple1 SET value=20 WHERE id=1; SECONDARY> UPDATE simple2 SET value=20 WHERE id=1; SECONDARY> COMMIT; SECONDARY> UPDATE simple3 SET value=20 WHERE id=1; # non conflicting SECONDARY> SELECT * FROM simple1; SELECT * FROM simple2; SELECT * FROM simple3; +----+-------+ | id | value | +----+-------+ | 1 | 20 | +----+-------+ +----+-------+ | id | value | +----+-------+ | 1 | 20 | +----+-------+ +----+-------+ | id | value | +----+-------+ | 1 | 20 | +----+-------+
If you now check the exception tables then you can see that the primary has received the changes from the secondary and because the first transaction updated the same row in simple1 during its window of conflict it has recorded that the change needs to be rolled back – this will happen as soon as the replication thread is restarted on the secondary:
PRIMARY> SELECT * FROM simple1$EX\G *************************** 1. row *************************** NDB$server_id: 9 NDB$master_server_id: 20 NDB$master_epoch: 31344671326216 NDB$count: 3 NDB$OP_TYPE: UPDATE_ROW NDB$CFT_CAUSE: TRANS_IN_CONFLICT NDB$ORIG_TRANSID: 193282978304 id: 1 value$OLD: 10 value$NEW: 20 1 row in set (0.00 sec) PRIMARY> SELECT * FROM simple2$EX\G *************************** 1. row *************************** NDB$server_id: 9 NDB$master_server_id: 20 NDB$master_epoch: 31344671326216 NDB$count: 1 NDB$OP_TYPE: UPDATE_ROW NDB$CFT_CAUSE: TRANS_IN_CONFLICT NDB$ORIG_TRANSID: 193282978304 id: 1 value$OLD: 10 value$NEW: 20 PRIMARY> SELECT * FROM simple3$EX\G Empty set (0.01 sec) SECONDARY> START SLAVE IO_THREAD; SECONDARY> SELECT * FROM simple1; SELECT * FROM simple2; SELECT * FROM simple3; +----+-------+ | id | value | +----+-------+ | 1 | 13 | +----+-------+ +----+-------+ | id | value | +----+-------+ | 1 | 10 | +----+-------+ +----+-------+ | id | value | +----+-------+ | 1 | 20 | +----+-------+
These are the results we expect – simple1 has the value set by the primary with the subsequent change on the secondary rolled back; simple2 was not updated by the primary but the change on the secondary was rolled back as it was made in the same transaction as the conflicting update to simple1. The change on the secondary to simple3 has survived as it was made outside of any conflicting transaction and the change was not dependent on any conflicting changes. Finally just confirm that the data is identical on the primary:
PRIMARY> SELECT * FROM simple1; SELECT * FROM simple2; SELECT * FROM simple3; +----+-------+ | id | value | +----+-------+ | 1 | 13 | +----+-------+ +----+-------+ | id | value | +----+-------+ | 1 | 10 | +----+-------+ +----+-------+ | id | value | +----+-------+ | 1 | 20 | +----+-------+
Statistics are provided on the primary that record that 1 conflict has been detected, effecting 1 transaction and that it resulted in 2 row changes being rolled back:
PRIMARY> SHOW STATUS LIKE 'ndb_conflict%'; +------------------------------------------+----------------+ | Variable_name | Value | +------------------------------------------+----------------+ | Ndb_conflict_fn_max | 0 | | Ndb_conflict_fn_old | 0 | | Ndb_conflict_fn_max_del_win | 0 | | Ndb_conflict_fn_epoch | 0 | | Ndb_conflict_fn_epoch_trans | 0 | | Ndb_conflict_fn_epoch2 | 0 | | Ndb_conflict_fn_epoch2_trans | 1 | | Ndb_conflict_trans_row_conflict_count | 1 | | Ndb_conflict_trans_row_reject_count | 2 | | Ndb_conflict_trans_reject_count | 1 | | Ndb_conflict_trans_detect_iter_count | 1 | | Ndb_conflict_trans_conflict_commit_count | 1 | | Ndb_conflict_epoch_delete_delete_count | 0 | | Ndb_conflict_reflected_op_prepare_count | 0 | | Ndb_conflict_reflected_op_discard_count | 0 | | Ndb_conflict_refresh_op_count | 0 | | Ndb_conflict_last_conflict_epoch | 37391985278995 | | Ndb_conflict_last_stable_epoch | 37417755082760 | +------------------------------------------+----------------+ SECONDARY> SHOW STATUS LIKE 'ndb_conflict%'; +------------------------------------------+----------------+ | Variable_name | Value | +------------------------------------------+----------------+ | Ndb_conflict_fn_max | 0 | | Ndb_conflict_fn_old | 0 | | Ndb_conflict_fn_max_del_win | 0 | | Ndb_conflict_fn_epoch | 0 | | Ndb_conflict_fn_epoch_trans | 0 | | Ndb_conflict_fn_epoch2 | 0 | | Ndb_conflict_fn_epoch2_trans | 0 | | Ndb_conflict_trans_row_conflict_count | 0 | | Ndb_conflict_trans_row_reject_count | 0 | | Ndb_conflict_trans_reject_count | 0 | | Ndb_conflict_trans_detect_iter_count | 0 | | Ndb_conflict_trans_conflict_commit_count | 0 | | Ndb_conflict_epoch_delete_delete_count | 0 | | Ndb_conflict_reflected_op_prepare_count | 1 | | Ndb_conflict_reflected_op_discard_count | 1 | | Ndb_conflict_refresh_op_count | 2 | | Ndb_conflict_last_conflict_epoch | 32280974196749 | | Ndb_conflict_last_stable_epoch | 18897856102416 | +------------------------------------------+----------------+
These status variables are described in the MySQL Cluster documentation.
You can restart MySQL Cluster processes (nodes) without losing database service (for example if adding extra memory to a server) and so on the face of it, the speed of the restarts isn’t that important. Having said that, while the node is restarting you’ve lost some of your high-availability which for super-critical applications can make you nervous. Additionally, faster restarts mean that you can complete maintenance activities faster – for example, a software upgrade requires a rolling restart of all of the nodes – if you have 48 data nodes then you want each of the data nodes to restart as quickly as possible.
MySQL 7.4 includes a number of optimisations to the restart code and so if you’re already using MySQL Cluster, it might be interesting to see how much faster it gets for your application.
MySQL Cluster presents a lot of monitoring information through the ndbinfo database and in 7.4 we’ve added some extra information on how memory is used for individual tables and how operations are distributed.
MySQL Cluster allocates all of the required memory when a data node starts and so any information on memory usage from the operating system is of limited use and provides no clues as to how memory is used with the data nodes – for example, which tables are using the most memory. Also, as this is a distributed database, it is helpful to understand whether a particular table is using a similar amount of memory in each data node (if not then it could be that a better partitioning/sharding key could be used). Finally, when rows are deleted from a table, the memory for those rows would typically remain allocated against that table and so it is helpful to understand how many of these ‘empty slots’ are available for use by new rows in that table. MySQL Cluster 7.4 introduces a new table – ndbinfo.memory_per_fragment – that provides that information.
For example; to see how much memory is being used by each data node for a particular table…
mysql> CREATE DATABASE clusterdb;USE clusterdb; mysql> CREATE TABLE simples (id INT NOT NULL AUTO_INCREMENT PRIMARY KEY) ENGINE=NDB; mysql> SELECT node_id AS node, fragment_num AS frag, \ fixed_elem_alloc_bytes alloc_bytes, \ fixed_elem_free_bytes AS free_bytes, \ fixed_elem_free_rows AS spare_rows \ FROM ndbinfo.memory_per_fragment \ WHERE fq_name LIKE '%simples%'; +------+------+-------------+------------+------------+ | node | frag | alloc_bytes | free_bytes | spare_rows | +------+------+-------------+------------+------------+ | 1 | 0 | 131072 | 5504 | 172 | | 1 | 2 | 131072 | 1280 | 40 | | 2 | 0 | 131072 | 5504 | 172 | | 2 | 2 | 131072 | 1280 | 40 | | 3 | 1 | 131072 | 3104 | 97 | | 3 | 3 | 131072 | 4256 | 133 | | 4 | 1 | 131072 | 3104 | 97 | | 4 | 3 | 131072 | 4256 | 133 | +------+------+-------------+------------+------------+
When you delete rows from a MySQL Cluster table, the memory is not actually freed up and so if you check the existing memoryusage table you won’t see a change. This memory will be reused when you add new rows to that same table. In MySQL Cluster 7.4, it’s possible to see how much memory is in that state for a table…
mysql> SELECT node_id AS node, fragment_num AS frag, \ fixed_elem_alloc_bytes alloc_bytes, \ fixed_elem_free_bytes AS free_bytes, \ fixed_elem_free_rows AS spare_rows \ FROM ndbinfo.memory_per_fragment \ WHERE fq_name LIKE '%simples%'; +------+------+-------------+------------+------------+ | node | frag | alloc_bytes | free_bytes | spare_rows | +------+------+-------------+------------+------------+ | 1 | 0 | 131072 | 5504 | 172 | | 1 | 2 | 131072 | 1280 | 40 | | 2 | 0 | 131072 | 5504 | 172 | | 2 | 2 | 131072 | 1280 | 40 | | 3 | 1 | 131072 | 3104 | 97 | | 3 | 3 | 131072 | 4256 | 133 | | 4 | 1 | 131072 | 3104 | 97 | | 4 | 3 | 131072 | 4256 | 133 | +------+------+-------------+------------+------------+ mysql> DELETE FROM clusterdb.simples LIMIT 1; mysql> SELECT node_id AS node, fragment_num AS frag, \ fixed_elem_alloc_bytes alloc_bytes, \ fixed_elem_free_bytes AS free_bytes, \ fixed_elem_free_rows AS spare_rows \ FROM ndbinfo.memory_per_fragment \ WHERE fq_name LIKE '%simples%'; +------+------+-------------+------------+------------+ | node | frag | alloc_bytes | free_bytes | spare_rows | +------+------+-------------+------------+------------+ | 1 | 0 | 131072 | 5504 | 172 | | 1 | 2 | 131072 | 1312 | 41 | | 2 | 0 | 131072 | 5504 | 172 | | 2 | 2 | 131072 | 1312 | 41 | | 3 | 1 | 131072 | 3104 | 97 | | 3 | 3 | 131072 | 4288 | 134 | | 4 | 1 | 131072 | 3104 | 97 | | 4 | 3 | 131072 | 4288 | 134 | +------+------+-------------+------------+------------+
As a final example, we can check whether a table is being evenly sharded accross the data nodes (in this case a realy bad sharding key was chosen)…
mysql> CREATE TABLE simples (id INT NOT NULL AUTO_INCREMENT, \ species VARCHAR(20) DEFAULT "Human", PRIMARY KEY(id, species)) engine=ndb PARTITION BY KEY(species); // Add some data mysql> SELECT node_id AS node, fragment_num AS frag, \ fixed_elem_alloc_bytes alloc_bytes, \ fixed_elem_free_bytes AS free_bytes, \ fixed_elem_free_rows AS spare_rows \ FROM ndbinfo.memory_per_fragment \ WHERE fq_name LIKE '%simples%'; +------+------+-------------+------------+------------+ | node | frag | alloc_bytes | free_bytes | spare_rows | +------+------+-------------+------------+------------+ | 1 | 0 | 0 | 0 | 0 | | 1 | 2 | 196608 | 11732 | 419 | | 2 | 0 | 0 | 0 | 0 | | 2 | 2 | 196608 | 11732 | 419 | | 3 | 1 | 0 | 0 | 0 | | 3 | 3 | 0 | 0 | 0 | | 4 | 1 | 0 | 0 | 0 | | 4 | 3 | 0 | 0 | 0 | +------+------+-------------+------------+------------+
To ensure that resources are being used effectively, it is very helpful to understand the how each table is being access (how frequently and for what types of operations). To support this, the ndbinfo. operations_per_fragment table is provided. For example, the data in this table would let you identify that a large number of full table scans are performed on a particular table.
It is also important to identify if there are any hotspots where a disproportionate share of the queries for a table are hitting a particular fragment/data node. Again, ndbinfo. operations_per_fragment provides this information.
As an example of how to use some of the data from this table, a simple table is created and populated and then ndbinfo.operations_per_fragment is used to monitor how many Primary Key reads and table scans are performed[1]:
mysql> CREATE TABLE simples (id INT AUTO_INCREMENT PRIMARY KEY, time TIMESTAMP) ENGINE=NDB; mysql> SELECT fq_name AS 'Table', node_id AS 'Data Node', tot_key_reads AS 'Reads', tot_frag_scans AS 'Scans' FROM ndbinfo.operations_per_fragment WHERE fq_name LIKE '%simples'; +-----------------------+-----------+-------+-------+ | Table | Data Node | Reads | Scans | +-----------------------+-----------+-------+-------+ | clusterdb/def/simples | 3 | 0 | 1 | | clusterdb/def/simples | 3 | 0 | 0 | | clusterdb/def/simples | 4 | 0 | 0 | | clusterdb/def/simples | 4 | 0 | 1 | +-----------------------+-----------+-------+-------+ mysql> REPLACE INTO simples VALUES (); # Repeated several times mysql> SELECT * FROM simples; +----+---------------------+ | id | time | +----+---------------------+ | 7 | 2015-01-22 15:12:42 | … | 8 | 2015-01-22 15:12:58 | +----+---------------------+ 23 rows in set (0.00 sec) mysql> SELECT fq_name AS 'Table', node_id AS 'Data Node', tot_key_reads AS 'Reads', tot_frag_scans AS 'Scans' FROM ndbinfo.operations_per_fragment WHERE fq_name LIKE '%simples'; +-----------------------+-----------+-------+-------+ | Table | Data Node | Reads | Scans | +-----------------------+-----------+-------+-------+ | clusterdb/def/simples | 3 | 0 | 2 | | clusterdb/def/simples | 3 | 0 | 0 | | clusterdb/def/simples | 4 | 0 | 0 | | clusterdb/def/simples | 4 | 0 | 2 | +-----------------------+-----------+-------+-------+ mysql> SELECT * FROM simples WHERE id=11; +----+---------------------+ | id | time | +----+---------------------+ | 11 | 2015-01-22 15:12:59 | +----+---------------------+ mysql> SELECT fq_name AS 'Table', node_id AS 'Data Node', tot_key_reads AS 'Reads', tot_frag_scans AS 'Scans' FROM ndbinfo.operations_per_fragment WHERE fq_name LIKE '%simples'; +-----------------------+-----------+-------+-------+ | Table | Data Node | Reads | Scans | +-----------------------+-----------+-------+-------+ | clusterdb/def/simples | 3 | 0 | 2 | | clusterdb/def/simples | 3 | 0 | 0 | | clusterdb/def/simples | 4 | 0 | 0 | | clusterdb/def/simples | 4 | 1 | 2 | +-----------------------+-----------+-------+-------+
Note that there are two rows listed for each data node but only one row for each has non-zero values; this is because each data node holds the primary fragment for one of the partitions and the secondary fragment for the other – all operations are performed only on the active fragments. This is made clearer if the fragment number is included in the query:
mysql> SELECT fq_name AS 'Table', node_id AS 'Data Node', fragment_num AS 'Fragment', tot_key_reads AS 'Reads', tot_frag_scans AS 'Scans' FROM ndbinfo.operations_per_fragment WHERE fq_name LIKE '%simples'; +-----------------------+-----------+----------+-------+-------+ | Table | Data Node | Fragment | Reads | Scans | +-----------------------+-----------+----------+-------+-------+ | clusterdb/def/simples | 3 | 0 | 0 | 2 | | clusterdb/def/simples | 3 | 1 | 0 | 0 | | clusterdb/def/simples | 4 | 0 | 0 | 0 | | clusterdb/def/simples | 4 | 1 | 1 | 2 | +-----------------------+-----------+----------+-------+-------+
We’re really excited about the GA for this new MySQL Cluster release; if you get chance to try it out then please let us know how you get on – either through a comment on this blog, a MySQL bug report or a post to the MySQL Cluster Forum.
Earlier this week, I presented a webinar on MySQL High Availability options for MySQL – what they are and how to choose the most appropriate one for your application.
The replay of this webinar can now be viewed here or if you just want to look at the charts then scroll down. At the end of this post, I include a summary of the Q&A from the webinar.
How important is your data? Can you afford to lose it? What about just some of it? What would be the impact if you couldn’t access it for a minute, an hour, a day or a week?
Different applications can have very different requirements for High Availability. Some need 100% data reliability with 24x7x365 read & write access while many others are better served by a simpler approach with more modest HA ambitions.
MySQL has an array of High Availability solutions ranging from simple backups, through replication and shared storage clustering – all the way up to 99.999% available shared nothing, geographically replicated clusters. These solutions also have different ‘bonus’ features such as full InnoDB compatibility, in-memory real-time performance, linear scalability and SQL & NoSQL APIs.
The purpose of this webinar is to help you decide where your application sits in terms of HA requirements and discover which of the MySQL solutions best fit the bill. It will also cover what you need outside of the database to ensure High Availability – state of the art monitoring being a prime example.
I’ll be presenting a webinar tomorrow (Thursday 19th February) on MySQL High Availability options for MySQL – what they are and how to choose the most appropriate one for your application.
As always, the webinar is free but you should register here. Even if you can’t attend, it’s worth registering as you’ll be sent a link to the replay as soon as it’s available.
How important is your data? Can you afford to lose it? What about just some of it? What would be the impact if you couldn’t access it for a minute, an hour, a day or a week?
Different applications can have very different requirements for High Availability. Some need 100% data reliability with 24x7x365 read & write access while many others are better served by a simpler approach with more modest HA ambitions.
MySQL has an array of High Availability solutions ranging from simple backups, through replication and shared storage clustering – all the way up to 99.999% available shared nothing, geographically replicated clusters. These solutions also have different ‘bonus’ features such as full InnoDB compatibility, in-memory real-time performance, linear scalability and SQL & NoSQL APIs.
The purpose of this webinar is to help you decide where your application sits in terms of HA requirements and discover which of the MySQL solutions best fit the bill. It will also cover what you need outside of the database to ensure High Availability – state of the art monitoring being a prime example.
WHEN: