NDB did not start if IPv6 support was not enabled on the host, even when no nodes in the cluster used any IPv6 addresses. (Bug #106485, Bug #33324817, Bug #33870642, WL #15661)

Important Change; NDB Cluster APIs: The NdbRecord interface allows equal changes of primary key values; that is, you can update a primary key value to its current value, or to a value which compares as equal according to the collation rules being used, without raising an error. NdbRecord does not itself try to prevent the update; instead, the data nodes check whether a primary key is updated to an unequal value and in this case reject the update with Error 897: Update attempt of primary key via ndbcluster internal api.

Previously, when using any other mechanism than NdbRecord in an attempt to update a primary key value, the NDB API returned error 4202 Set value on tuple key attribute is not allowed, even setting a value identical to the existing one. With this release, the check when performing updates by other means is now passed off to the data nodes, as it is already by NdbRecord.

This change applies to performing primary key updates with NdbOperation::setValue(), NdbInterpretedCode::write_attr(), and other methods of these two classes which set column values (including NdbOperation methods incValue(), subValue(), NdbInterpretedCode methods add_val(), sub_val(), and so on), as well as the OperationOptions::OO_SETVALUE extension to the NdbOperation interface. (Bug #35106292)

NDB Cluster APIs: Printing of debug log messages was enabled by default for Ndb_cluster_connection. (Bug #35416908)

References: See also: Bug #35927.

NDB Cluster APIs: While setting up an NdbEventOperation, it is possible to pass a pointer to a buffer provided by the application; when data is later received, it should be available in that specified location.

The received data was properly placed in the provided buffer location, but the NDB API also allocated internal buffers which, subsequently, were not actually needed, ultimately wasting resources. This problem primarily manifested itself in applications subscribing to data changes from NDB using the NdbEventOperation::getValue() and getPreValue() functions with the buffer provided by application.

To remedy this issue, we no longer allocate internal buffers in such cases. (Bug #35292716)

When dropping an NdbEventOperation after use, the ndbcluster plugin now first explicitly clears the object's custom data area. (Bug #35424845)

After a socket polled as readable in NdbSocket::readln(), it was possible for SSL_peek() to block in the kernel when the TLS layer held no application data. We fix this by releasing the lock on the user mutex during SSL_peek(), as well as when polling. (Bug #35407354)

When handling the connection (or reconnection) of an API node, it was possible for data nodes to inform the API node that it was permitted to send requests too quickly, which could result in requests not being delivered and subsequently timing out on the API node with errors such as Error 4008 Receive from Ndb failed or Error 4012 Request ndbd time-out, maybe due to high load or communication problems. (Bug #35387076)

Made the following improvements in warning output:

(Bug #35376705)

Memory consumption of long-lived threads running inside the ndbcluster plugin grew when accessing the data dictionary. (Bug #35362906)

A failure to connect could lead ndb_restore to exit with code 1, without reporting any error message. Now we supply an appropriate error message in such cases. (Bug #35306351)

When deferred triggers remained pending for an uncommitted transaction, a subsequent transaction could waste resources performing unnecessary checks for deferred triggers; this could lead to an unplanned shutdown of the data node if the latter transaction had no committable operations.

This was because, in some cases, the control state was not reinitialized for management objects used by DBTC.

We fix this by making sure that state initialization is performed for any such object before it is used. (Bug #35256375)

A pushdown join between queries featuring very large and possibly overlapping IN() and NOT IN() lists caused SQL nodes to exit unexpectedly. One or more of the IN() (or NOT IN()) operators required in excess of 2500 arguments to trigger this issue. (Bug #35185670, Bug #35293781)

The buffers allocated for a key of size MAX_KEY_SIZE were of insufficient size. (Bug #35155005)

The fix for a previous issue added a check to ensure that fragmented signals are never sent to V_QUERY blocks, but this check did not take into account that, when the receiving node is not a data node, the block number is not applicable. (Bug #35154637)

References: This issue is a regression of: Bug #34776970.

ndbcluster plugin log messages now use SYSTEM as the log level and NDB as the subsystem for logging. This means that informational messages from the ndbcluster plugin are always printed; their verbosity can be controlled by using --ndb_extra_logging. (Bug #35150213)

We no longer print an informational message Validating excluded objects to the SQL node's error log every ndb_metadata_check_interval seconds (default 60) when log_error_verbosity is greater than or equal to 3 (INFO level). It was found that such messages flooded the error log, making it difficult to examine and using excess disk space, while not providing any additional benefit. (Bug #35103991)

Some calls made by the ndbcluster handler to push_warning_printf() used severity level ERROR, which caused an assertion in debug builds. This fix changes all such calls to use severity WARNING instead. (Bug #35092279)

When a connection between a data node and an API or management node was established but communication was available only from the other node to the data node, the data node considered the other node “live”, since it was receiving heartbeats, but the other node did not monitor heartbeats and so reported no problems with the connection. This meant that the data node assumed wrongly that the other node was (fully) connected.

We solve this issue by having the API or management node side begin to monitor data node liveness even before receiving the first REGCONF signal from it; the other node sends a REGREQ signal every 100 milliseconds, and only if it receives no REGCONF from the data node in response within 60 seconds is the node reported as disconnected. (Bug #35031303)

The log contained a high volume of messages having the form DICT: index index number stats auto-update requested, logged by the DBDICT block each time it received a report from DBTUX requesting an update. These requests often occur in quick succession during writes to the table, with the additional possibility in this case that duplicate requests for updates to the same index were being logged.

Now we log such messages just before DBDICT actually performs the calculation. This removes duplicate messages and spaces out messages related to different indexes. Additional debug log messages are also introduced by this fix, to improve visibility of the decisions taken and calculations performed. (Bug #34760437)

A comparison check in Dblqh::handle_nr_copy() for the case where two keys were not binary-identical could still compare as equal by collation rules if the key had any character columns, but did not actually check for the existence of the keys. This meant it was possible to call xfrm_key() with an undefined key. (Bug #34734627)

References: See also: Bug #34681439. This issue is a regression of: Bug #30884622.

Local checkpoints (LCPs) wait for a global checkpoint (GCP) to finish for a fixed time during the end phase, so they were performed sometimes even before all nodes were started.

In addition, this bound, calculated by the GCP coordinator, was available only on the coordinator itself, and only when the node had been started (start phase 101).

These two issues are fixed by calculating the bound earlier in start phase 4; GCP participants also calculate the bound whenever a node joins or leaves the cluster. (Bug #32528899)