In previous releases, frequent client connection attempts could result in excessive messages in the log file, although the messages were meant to be limited to one every 60 seconds. This issue has been resolved and the rapidly repeated messages are now muted.

There was an edge case where a voltadmin dr reset command could result in a deadlock, causing the database to hang. The issue has been resolved.

jQuery V3.5.1

jQuery UI V1.12.1

jQuery Slimscroll V1.3.8

jQuery Validate V1.19.2

There was an issue where, if a stored procedure queued more than 200 SQL statements before calling voltExecuteSQL() and at least one of the statements was a SELECT statement that returned data, the result buffer could become corrupted causing one or more nodes to crash. This issue has been resolved.

There was a problem where, if a properties file in the database root was corrupted, the database would issue a fatal error with no explanation. The error now identifies the corrupted file and the names of the missing properties.

In the situation where a cluster failed or was forcibly shutdown while a node was being added or removed, attempting to restart the cluster could result in an error claiming there were "incomplete command logs", even if command logging was not enabled. This was caused by an incomplete snapshot left by the interrupted cluster expansion. The issue has been resolved.

Additional information is now logged if the SQL compiler encounters an unexpected error while processing a data definition language (DDL) statement.

There was a rare condition where the VoltDB network process could report an index out of bounds error, causing the cluster to hang. This condition is now caught. As a consequence of the error, one of the nodes will stop, but the cluster as a whole will continue and not be deadlocked.

There was an issue where using the CAST function to convert a VARCHAR column to a BIGINT could generate incorrect values if the number in question had more than 18 digits. This issue has been resolved.

Support for a new, improved license format

The voltadmin show license command for displaying information about the current license for a running VoltDB cluster

A new selector, LICENSE, for the @SystemInformation system procedure to provide similar information programmatically

A new voltadmin inspect command used by VoltDB product support to display summary information about the cluster operating environment

Previously, if the snapshot rate limit was set (using the Java property SNAPSHOT_RATELIMIT_MEGABYTES), requesting a CSV formatted snapshot could raise an illegal argument exception stating that "requested permits must be positive" and the resulting snapshot files would be empty. This only affected CSV formatted snapshots. This problem has been resolved.

There was a potential situation where, if a cluster used for cross datacenter replication (XDCR) suffered one or more node failures, then was shutdown and restarted using command logs to recover, replication might later fail with a "replica ahead of master" error. This underlying issue was related to recovery using the failed node's command logs which did not match the current state of the remote cluster. This problem has been resolved.

A similar problem could occur if a node becomes detached from the cluster (for example, due to network issues) and does not immediately fail but times out. The result was that the remote cluster might stop replication, also reporting the "replica ahead of master" error. This issue has also been resolved.

There was a rare condition where workloads containing lots of complex multi-partition read transactions interspersed with single-partition writes could result in a deadlock, stalling the database and generating log messages reporting a "possible multipartition transaction deadlock". This issue has been resolved.

The snapshotconverter utility lets you generate CSV files from VoltDB snapshot files. These files can be used to recover and reload data from individual tables through the csvloader utility. However, for certain data — such as XDCR tables, tables defined with MIGRATE, or views with no COUNT(*) column — the snapshotconverter utility includes hidden columns in its output, which can be confusing. A new command flag has been added, --filter-hidden, that lets you exclude these hidden columns from the utility's output.

There was an issue where a stream could stop writing data to its export target after having more than two billion rows inserted into any one partition. The problem surfaced only after the necessary number of records (approximately 2.15 billion) were written to the export connector and the database was saved, shutdown, restarted, and restored. After the snapshot was restored, no further records were written to the target by the export connector.

This issue has now been resolved. In fact, upgrading VoltDB using the standard voltadmin shutdown --save command, installing this release, and then restarting the database will automatically circumvent the issue.

Due to issues in the underlying library used, it was possible for the JSON functions to return results in a different order on different servers, causing a hash mismatch error. This inconsistency, and the resulting issue, have now be resolved.

The VoltDB bulk loader (available in the client API and used in the loader utilities such as csvloader) has been optimized to remove an unnecessary regular expression evaluation of string columns. This change produces a noticeable improvement in load times for large data sets.

The snapshotconvert utility has been corrected to interpret null values as end-of-file, rather than reporting an error. At the same time, general error handling has been enhanced and extended to report more detailed information when a failure occurs.

There was a problem with the Kinesis importer where the importer could fail with a "no class found" error. This issue has been resolved.

There was a rare situation where if a schema failed causing a deadlock, subsequent attempts to rejoin nodes to the cluster would fail. This issue has been resolved.

Previously, the bulkloader interface, which is used by VoltDB data utilities such as csvloader and is available through the Java API, did not correctly account for the additional data structures required by cross datacenter replication (XDCR) or TTL with migrate. As a result, attempting to bulk load data into an XDCR cluster or a table with MIGRATE TO TARGET and USING TTL could cause the cluster to crash. This issue has been resolved.

There was an issue with earlier releases of the JDBC export connector. If the connector property createtable was set to true but the target database (for example, Oracle) did not support the CREATE TABLE... IF NOT EXISTS clause, the export connector would repeatedly attempt to create the table and fail, although the table did exist, resulting in innumerable spurious error messages. This issue has been resolved.

In earlier releases, if an attempt was made to restore a snapshot to an empty database, but the specified unique ID did not exist or was misspelled, subsequent attempts to restore the snapshot would claim to succeed but would not restore any data. This issue has been resolved.

Previously, if two clusters established cross datacenter replication (XDCR), including a table defined with USING TTL and MIGRATE TO TARGET on one cluster but not on the other, the cluster could crash with a fatal error while processing the XDCR binary logs. This issue has been resolved.

There was a rare condition where the digest of the synchronization snapshot for database replication (DR or XDCR) could exceed the 50MB limit for network packets. When this happened, nodes on the consumer cluster would fail. The size of the digest files triggering this condition have been reduced to eliminate the issue.

There was an issue where a schema or configuration change on a running database could stall if there was an active export queue connected to a JDBC target. Closing the export connection took too long and could potentially cause a deadlock with the requested change. This issue has been resolved.

There was a rare case where, if a replica cluster in passive DR is promoted and then a node fails, other nodes could subsequently fail when the cluster tries to reassign partition leadership. The symptom for this particular failure scenario was that the failing nodes reported an error processing an invocation dispatcher request. This issue has now been resolved.

Previously, you could not enable SSL/TLS and Kerberos authentication at the same time. This limitation has been removed.

Using VoltDB V6.5, taking a snapshot of a cluster with database replication (DR) enabled could create a snapshot that cannot be restored in later versions of VoltDB (V7, V8 or V9). The restore operation would fail with an error. This issue has been resolved and the current version of VoltDB can now restore the problematic snapshots.

There was an issue where restoring a snapshot, either manually or automatically on startup, could fail if the directory contained multiple snapshots and the unique IDs for the snapshots were similar. (That is, one ID matched the starting characters of another ID, such as test and testme.) When this happened, the restore would fail with an error stating that a table had an "inconsistent transaction ID." This issue has been resolved. VoltDB now performs an exact match on the selected snapshot's unique identifier.

Previously, attempting to restore a snapshot created on a standalone cluster to a cluster configured for cross datacenter replication (XDCR) would fail with a misleading error message (indicating that the configuration could not be updated). This issue has been resolved, the snapshot is restored, and no error is reported.

There was an obscure issue where if a database cluster was restored from snapshots multiple times, the fourth restore command could stall and eventually timeout. For this unusual situation to occur the database schema must contain views, must be K safe, and nodes must have stopped and rejoined between each snapshot restore operation. This issue has been resolved.

When manually restoring a snapshot, you must specify a path to the snapshot files. Previously, if the path did not exist on a node, the restore would fail, even if sufficient files for a full backup existed on the other nodes. The restore now ignores the missing path.

There was a problem where using the VoltDB Management Center to restore a snapshot would not work. Using the Restore button on the web interface would actually start two restore operations, which would then cause constraint violations. This issue has been resolved.

When rejoining a node to a running cluster , the system clock on the rejoining node must be within the limits for clock skew on the cluster, just like when starting the cluster for the first time. If not, the rejoin operation will fail. Previously, there was an issue where if a rejoin failed due to clock skew, subsequent attempts to rejoin nodes would fail even if the clock skew had been corrected. This issue has been resolved.

VoltDB 8.2 introduced an issue that could cause a cluster to hang or crash when attempting to add nodes using elastic scaling. Under certain conditions, where a database has views, it was possible for the cluster to hang or to crash reporting a deserialization error while attempting to pass copies of the current database contents to the joining nodes. This issue has been resolved.

Previously, while adding nodes to the cluster on the fly, if a table exceeded its LIMIT PARTITION ROWS setting, the elastic expansion of the cluster would fail along with one of the existing nodes. This issue has been resolved.

There was an issue where, if a database had no persistent tables (but may have had streams defined with or without export), a node rejoining the cluster could end up with inconsistent internal settings related to export and/or database replication. This issue has been resolved.

Under certain rare error conditions, an attempt to elastically add nodes to a cluster could fail, crashing the cluster and reporting an illegal argument exception. This issue has been resolved.

There was an issue with elastic expansion of a cluster. If, while adding nodes "on the fly", ongoing transactions within one of the partitions being rebalanced generates a constraint violation, it could result in the rebalance operation reporting a fatal "failed to delete tuple" error, causing the server process to exit. This issue has now been resolved.

In previous releases, attempting to rejoin a node to a cluster could fail with the error "Bad message length exception" if the database schema contained too many tables (that is, thousands of table declarations, regardless of actual data volume). This issue has now been resolved.

The New Relic latency graph data has been adjusted to improve accuracy.

The New Relic node count report was accurate while the cluster was running, but did not "zero out" for periods while the cluster was down. This report now reports zero running nodes for those intervals when the cluster was stopped.

There was an issue where changing the Kafka export configuration on a running database could cause the update to hang. The issue was triggered by a Kafka export configuration specifying an invalid Kafka broker. Attempting to update the configuration to change the invalid broker specification would cause the update process to hang. This issue has been resolved.

VoltDB V8.4.9 introduced an obscure but critical bug where a SQL query with a complex ORDER BY clause and two separate column expressions each using the DECODE() function could produce incorrect results. This issue has been resolved.

Previously, if a GROUP BY query included an arithmetic expression combining an aggregate function and a parameter cast to a specific datatype with the CAST() function, the query failed to compile. For example:

SELECT a, CAST(? AS INTEGER) + COUNT(b) 
   FROM TABLE c GROUP BY a;

This issue has been resolved for the CAST() function. However, combining other parameterized functions with an aggregate may also cause a compilation error. The workaround, until the more general case is fixed, is to put the parameterized function in a subquery.

Previously, there was a rare case where an ad hoc query using parameters in a complex combination of CAST() and SUM() operations would give different results than the same query using constant values. For example, the following fragment with the arguments "A" and 1 gave different results than expected if the placeholders were replaced by the SQL constants 'A' and 1:

SUM( DECODE(column1, CAST(? AS VARCHAR), CAST(? AS INTEGER), NULL))

This issue has been resolved.

There was an issue where SNMP traps would fail if the server had been running for around 24 days. After 24 days, whenever an SNMP trap was triggered, VoltDB would report that the trap failed with an illegal argument exception error. This issue has been resolved.

All SNMP traps are now recorded in the log file. Previously, SNMP traps were sent, but not recorded. Log entries provide additional confirmation that SNMP events are triggered.

There was an edge case where the database could stop processing snapshots if there were too many tables with too much data. The issue involved a race condition and did not surface all the time, so the database might run for an extended period of time before an error was triggered. But the more tables or the larger volume of data in the database, the more likely a snapshot might fail. And once it failed, no subsequent snapshots could be taken. This issue has been resolved.

As part of the overall hardening of the LTS release, a number of rare circumstances causing deadlocks were identified, tracked down, and resolved. These corner cases included a deadlock during a node rejoin, when a node failed while processing a read-only multi-partition procedure, and as a result of nodes failing while the cluster is trying to re-establish a quorum from an earlier node failure. All these issues have now been resolved.

Commons-compress 1.19

Dom4j 2.1.1

Jackson 2.9.10

Jetty 3.27

Kafka client 0.10.2.2

Netty 4.1.43

Openssl 1.0.2t

Scala 2.11.12

Tomcat 7.0.96

If a node rejoining a cluster was stopped with @StopNode but later successfully rejoined the cluster, there was a chance the other nodes in the cluster did not recognize the successful rejoin and would continue reporting the error "REJOIN: No stream snapshot ack message was received in the past 10 minutes or the thread was interrupted." This issue is now resolved.

Under certain rare conditions, if a node failed on a K-safe cluster while a node was being added, a snapshot was in progress, or the @SwapTables system procedure was executing, the command logs for that cluster could be rendered incomplete. If, after this event, the cluster stopped and restarted before the next command log snapshot was taken, the command logs could not be replayed beyond the point of the node failure. This issue has now been resolved.

There was an issue with file export when using TSV (tab-separated value) format and exporting data containing quotations marks or backslashes. The export connector incorrectly attempted to "escape" the output, although TSV format does not support quoting or escaping. The result was incorrect output in the export file. This issue has been resolved. The export connector no longer attempts to escape special characters in the output.

Export data optionally contains six columns of metadata, including a timestamp identifying when the row was exported. Previously, this timestamp was mistakenly set 12 years prior to the actual date. This issue has been resolved.

This release contains a substantial number of fixes and improvements to the handling of export data. In particular, it addresses several edge cases associated with the durability of export queues during unusual system failure scenarios. Taken together these changes significantly improve the reliability, durability, and recoverability of export in VoltDB.

There was issue where, in rare cases, if a schema update or configuration change failed, subsequent attempts to update the schema or configuration would also fail. This could only happen if the original update failed with an unhandled exception (such as a reference to a missing Java method), at which point subsequent update attempts reported that another update was still in progress. This issue has been resolved.

There was an edge case where, if a multi-partition transaction encounters a deadlock during a rejoin, the rejoin — and any subsequent rejoins — would fail repeatedly reporting a possible deadlock. This issue has been resolved.

There was a rare edge case where a query could return incomplete results. If an index on a table included two columns and the WHERE clause of a query on the table included both an IN clause applied to one column and a less than or equal to (<=) evaluation of the second column, fewer than expected rows were selected. This issue has been resolved.

There was an issue with elastic expansion of a cluster. If, while adding nodes "on the fly", ongoing transactions within one of the partitions being rebalanced generates a constraint violation, it could result in the rebalance operation reporting a fatal "failed to delete tuple" error, causing the server process to exit. This issue has now been resolved.

There was a edge case where a K-safe cluster could run out of memory, caused by too many long-running and/or failing multi-partition transactions. In the case of K-safe clusters, the multi-part coordinator must maintain a log of its in-flight transactions. If too many transactions linger or fail and rollback, the log itself can grow out of control, ultimately exceeding available memory before it can be pruned and causing the server process to fail. This rare condition is now avoided by increasing the priority of log management, ensuring it is pruned in a timely fashion.

There was an obscure issue where if a database cluster was restored from snapshots multiple times, the fourth restore command could stall and eventually timeout. For this unusual situation to occur the database schema must contain views, must be K safe, and nodes must have stopped and rejoined between each snapshot restore operation. This issue has been resolved.

There was an issue in the VoltDB Enterprise Edition where, if you attempted to add a node on the fly (known as elastic scaling) but did not have a license for database replication (DR), the operation would cause the cluster to crash and interfere with restarting the cluster from command logs. This issue has been resolved.

There was an issue where the export subsystem was repeatedly reporting a warning that it received an export message with a signature "which does not exist on this node." Although this message was innocuous and did not indicate any real problem, it was misleading and could fill up the logs for no reason. This issue has now been resolved.

There was an edge case where attempting to update the schema while a snapshot is being saved could hang the database. This issue has been resolved. Attempts to change the schema or configuration during a snapshot are no longer allowed and must wait until the snapshot is complete.

There was an issue introduced in VoltDB 8.2 when the USING TTL clause was added to allow you to automatically delete old records from tables based on the specified column value. Accidentally the USING TTL clause was also allowed on CREATE STREAM statements, although it has no application to streams. This issue has now been resolved.

Under certain rare conditions, the @Quiesce system procedure could return control to the calling program before all export and DR data is successfully processed. If this occurs it was possible for an orderly shutdown (that is, voltadmin shutdown without the --force argument) to stop the cluster before all pending DR or export data was made durable. This rare race condition has now been resolved.

Errors and stack traces appearing during startup have been eliminated.

Files written to the database root directory during startup are now protected against corruption due to nodes failing during the startup process.

There was an issue in database replication (DR) where dropping a partitioned table could cause the consumer cluster to hang. This could only happen if the cluster was not paused and DR drained before the schema change and there was a simultaneous transaction being processed involving the table and other non-partitioned tables. This issue has now been resolved.

ASSUMEUNIQUE indexes are designed specifically for partitioned tables, where a partitioned procedure cannot verify the uniqueness of an index on columns other than the partitioning column. They do not apply to replicated tables. However, there was an issue where defining an ASSUMEUNIQUE index on a replicated table could result in a database failure. This issue has been resolved.

As part of this fix, use of CREATE ASSUMEUNIQUE INDEX on a replicated table will automatically be converted to CREATE UNIQUE INDEX. So to create an ASSUMEUNIQUE index on a partitioned table you must partition the table before issuing the CREATE ASSUMEUNIQUE INDEX statement or define the index as a constraint in the CREATE TABLE statement itself.

Similarly, declaring an index that should be ASSUMEUNIQUE (because it is on a non-partitioning column of a partitioned table) as UNIQUE could cause problems in previous releases. VoltDB correctly warned that the declaration of the UNIQUE index was invalid. But a later transaction could result in an "invalid use of UNIQUE" error, stopping the database. This issue has been resolved.

The LOG10() function can result in invalid values, depending on the argument to the function. Normally, using such functions to define an index on a table that already contains data is not allowed. However, LOG10() was inadvertently allowed in this situation. This issue has been resolved and use of LOG10() is no longer allowed when creating indexes on non-empty tables.

Recent changes to the export connector programming interface incorrectly changed the behavior of the onBlockStart() method. For custom export clients, onBlockStart() should be invoked only once per block. However, starting in VoltDB 8.1, multiple calls to onBlockStart() could be initiated before onBlockCompletion() . This change in behavior has been corrected.

There is a Java environment variable that is intended to set the maximum time interval between when data is inserted into a stream and when it is flushed to the export connector. However, in previous releases, setting that variable, MAX_EXPORT_BUFFER_FLUSH_INTERVAL, did not have any affect. This issue has now been resolved. You can now set the export flush interval in milliseconds, with 1000 being the current lowest practical value.

Errors when draining export — a number of issues related to the last few records in an export buffer not being able to drain have been resolved.

Export Statistics — The @Statistics system procedure has a new selector, EXPORT, that provides detailed information about the progress of export for each stream and partition.

Gaps in export queues — Exporting data from VoltDB is an asynchronous process, so that export does not impact the performance of the database itself if there are delays communicating with the external target. If the remote system cannot keep up, VoltDB queues the overflow data to disk. This both buffers the pending content and provides durability in case of database failure.

However, export data is only durable if the database servers restart with the same disk-based content. Similarly, in a K-safe environment, all copies of a partition buffer export data. But when a node stops and rejoins the cluster, there can be a gap in its export buffer for the time it is down. If the export target is not accepting data fast enough and the cluster nodes stop and rejoin frequently, it is possible for a gap to occur in the currently active export queue.

In that past, it was assumed the data was lost and export proceeded to the next available record. The database now handles export gaps more intelligently. First it queries all copies of the partition to see if any contain the missing records. If the records are not found but nodes are currently missing from the cluster, export waits for the nodes to rejoin to see if they have the missing export records. If so, export continues where it left off. If not and the cluster is complete, export resumes, logging the fact that certain export records may be lost. If, for any reason, export is paused at a gap in the queue and you wish to continue (for example, you cannot rejoin a failed node but want export to continue) you can use the voltadmin export release command to have export resume at the next available record.

Less disruption to ongoing transactions

No lost connection errors returned to client applications

Better management of export buffers, avoiding gaps in the pending queues

Enabling SSL encryption on VoltDB interfaces now takes advantage of OpenSSL to significantly improve performance if it is installed on the server. For the improvement to take effect, both the VoltDB server and Java client library must be upgraded to the latest version. The new client library also has a dependency on the Netty libraries io.netty:netty-all:4.1.32.Final and io.netty:netty-tcnative-boringssl-static:2.0.20.Final. If OpenSSL is not installed or the latest (8.3.4 or later) VoltDB client library used, the server falls back to using the builtin Java SSL client.

There was an issue with database replication where replication could stop and report that the replica was "ahead" of the master. This issue has been resolved.

Previously, VoltDB did not allow you to create an index using a unary minus. That is, a column expression preceded by a minus sign. For example "CREATE INDEX myindex ON mytable (product, -price)". This issue has been resolved.

In a related issue, VoltDB previously allowed DESC (descending) in the index definition, although it had no effect on the index created. The DESC keyword is no longer allowed.

There was an issue in the Java client API where an idle client application (that is, an application that creates a client instance but makes no calls to VoltDB for an extended period of time) could eventually use up all available CPU. This issue has been resolved.

In database replication (DR), if a transaction executed a TRUNCATE TABLE statement followed by inserts, the inserted data was present on the producer cluster but was not properly replicated to the consumer. This issue has been resolved.

VoltDB introduced two major changes in recent releases: significant improvement in replicated table storage (V8.1) and leadership rebalancing (V8.3). One outcome of these changes is the discovery of edge cases where multi-partition transactions stall in what is referred to as a "deadlock". These deadlocks are rare — often only seen in lab tests. Rare as they are, any such case of a processing failure is critical and this release fixes two known instances of multi-part deadlocks.

There was an issue introduced in 8.1 where, when using export and nodes failed and then rejoined the cluster, export could lose track of a few records, causing the export connector to never completely drain its queue. Symptoms of this bug were that the export statistics would never reach zero and attempts to use the voltadmin commands pause --wait or shutdown --save would hang. This issue has been resolved.

There was an issue where attempting to apply a schema change involving the "time to live" (TTL) feature through the sqlcmd utility could result in the command hanging. The root cause was an error applying the schema change. However, the error was not reported to the user. The sqlcmd utility now recognizes such situations, reports the error to the user, and returns to the command prompt. At the same time, the underlying issue with the TTL schema change has been corrected.

There was an issue introduced in 8.3 that interfered with database replication (DR) of multi-partition transactions. If the last query executed in a read/write transaction was a SELECT statement rather than a data manipulation statement (such as INSERT, UPDATE, or DELETE), the transaction might not be correctly replicated to the consumer cluster. This issue has been resolved.

The syntax of the ALTER TABLE for adding a USING TTL statement to an existing table was incorrect in 8.3. The correct syntax is ALTER TABLE table-name ADD USING TTL... but the "ADD" keyword was missing previously. This issue has been resolved.

Due to a change in the server logic for the VoltDB Management Center (VMC) introduced in V7.9, the database replication (DR) and Import tabs no longer showed up when those features were turned on in the database. This issue has been resolved.

The JSON interface no longer accepts HTTP requests that use the TRACE method.

There is a condition where, under certain conditions, if the database is idle (that is, no read or write transactions are occurring) snapshots can get into a scheduling loop causing a CPU spike and preventing other threads from running. This occurs only when the database is configured with a large number of sites per host running on systems with slower disks and fewer CPU cores (for example, in virtualized environments). To avoid this condition, a new option, DISABLE_​IMMEDIATE_​SNAPSHOT_​RESCHEDULING, has been added.

In normal database operation, this option is not needed. However, if you configuration matches these conditions and your database falls idle for any significant time, you can set this option to true when you start the database to circumvent the problem. You set the option as a Java environment variable on all the servers at start up using the VOLTDB_OPTS environment variable and including the "-D" flag. For example:

export VOLTDB_OPTS="-DDISABLE_IMMEDIATE_SNAPSHOT_RESCHEDULING=true"

Previously, if you initialized a new database root directory with a configuration that enabled cross datacenter replication (XDCR), it was not possible to restore a snapshot after starting the new, empty database. The problem was that XDCR creates streams for logging XDCR conflicts and those streams were seen as an existing schema. This issue has been resolved and XDCR conflict streams are ignored for the purposes of snapshot restore.

There was an issue introduced in 8.2 that affects database replication (DR) and time to live (TTL). If a replicated DR table is defined with TTL, the TTL delete procedure continuously generates DR binary logs. As a result, any attempt to pause the DR consumer or perform an orderly shutdown fails since the DR buffer never drains. Note that this issue was specific to databases with replicated tables declared as both DR tables and with TTL. This issue has now been resolved.

There was an issue in VoltDB 8.2 with the elastic expansion of clusters. After adding nodes to the cluster, it was possible for the rebalancing of the partitions to fail. This could only happen on VoltDB 8.2 or later when the schema includes a view on a partitioned table involved in the rebalancing. The failure could result in memory corruption, potentially leading to the cluster crashing at some future point in time. This issue has been resolved.

There was an extremely rare case where in a K-safe cluster, if a node failed and the remaining nodes are busy, the intra-cluster failure and repair messages could be processed out of order, causing errors in synchronization. The major difficulty with this rare error condition is that it would not be detected or reported by the cluster at the time. However, it could result in later transactions generating a hash mismatch. The issue was exacerbated with SSL/TLS enabled on the internal ports, since there could be a backlog of messages from the failed host requiring decryption before being delivered. This issue has been resolved.

When loading data into variable-width columns with the bulkloader methods (either through a utility such as csvloader or in a custom Java application), there was a memory leak that occurred whenever the input data exceeded the maximum size of the column. This issue was specific to VARCHAR and VARBINARY columns larger than 64 bytes and only affects VoltDB V8.1 or later. In extreme cases, where many such exceptions occurred, the process could potentially run out of memory. This issue has been resolved.

There was an issue with time to live (TTL) introduced in 8.2, where if too many records scheduled to be deleted had the exact same value for the TTL column, TTL would fail to delete the records and report an unexpected error condition. This issue has been resolved.

There was an issue introduced in VoltDB 8.1 related to database replication (DR) and TRUNCATE TABLE statements applied to replicated DR tables. Use of TRUNCATE TABLE on a producer cluster to clear a replicated table could result in memory corruption and failure of the consumer cluster. This issue has been resolved.

It should be noted that use of TRUNCATE TABLE on any tables in an XDCR environment (as opposed to passive DR) is not recommended, even with this issue resolved. The current implementation cannot guarantee that the two XDCR clusters might not suffer an undetectable conflict since a TRUNCATE TABLE statement does not log the specific rows that are deleted.

Normally, if the size of any input value exceeds the size of a variable-width column (such as VARCHAR or VARBINARY), it generates a SQL exception and the statement is rejected. There was an issue introduced in V8.1 where, under certain circumstances, when bulk loading data and the input exceeded the width of a VARCHAR or VARBINARY column, a fatal exception was generated and the database stopped. This did not happen during normal INSERT or UPDATE statements; only when using utilities such as csvloader or using the bulkloader methods in the Java client API. It also only occurred in certain operating system-specific environments. This issue has been resolved.

Due to a change in the server logic for the VoltDB Management Center (VMC) introduced in V7.9, the database replication (DR) and Import tabs no longer showed up when those features were turned on in the database. This issue has been resolved.

A query can have 1,025 parameters (or placeholders) at most. However, there was an issue where entering a query with too many parameters through the @AdHoc system procedure could result in a runtime error, crashing the database server. This includes queries entered through JDBC prepared statements, which use @AdHoc implicitly. This issue has been resolved.

Ability to "hide" the username and password for command line utilities in a separate credentials file. (See description above.)

When enabling SSL with a user-generated certificate, you need to specify both a keystore and truststore. When using a commercial certificate a local truststore should not be needed. However, previously VoltDB still required one. Specifying a truststore is no longer required when using a commercial certificate.

There was an issue with database replication (DR) where a large multi-partition transaction could produce binary logs from each partition on the producer that fell under the 50MB limit on inter-cluster communication. But when aggregated on the consumer for replay, the transaction exceeds the limit. The symptom was that nodes on the cluster would report a "bad message length" exception, causing nodes to be expelled from the cluster until the cluster itself failed. The possibility of an excessively large transaction still exists, but now the producer cluster rejects the transaction and replication and consistency between the clusters is maintained.

There was an issue using the VoltDB Management Center (VMC) if the VoltDB http port was set to port 80. Port 80 is the default port for web browsers and if the browser did not send a port number VMC would incorrectly assume a default of 8080 and not operate properly. This issue has been resolved.

VoltDB reserves the maximum negative value of numeric datatypes as null. (For example, -128 for a TINY integer.) So users should not be allowed to use these reserved values in the context of a given datatype. However, previously the compiler silently accepted such constants and interpreted them as null. This issue has been resolved. The compiler now throws an error when evaluating numeric values equal to a datatype's maximum negative value.

Certain SQL functions (such as NOW and PI) that take no argument can be entered with or without parentheses. However, these functions were not interpreted consistently in the selection list of a SELECT statement. If the parentheses were left off, NOW was interpreted as the function and PI was interpreted as a column reference. These functions are now both interpreted as functions, whether with or without parentheses.

Note: this is a slight change of behavior. If you use a column with the name PI, you will now have to fully qualify the name to have it be interpreted as a column rather than a function. For example, in the following statement, the first item in the section list is interpreted as the function PI and the second as the column PI:

SELECT pi, a.pi FROM a WHERE...

There was a rare condition involving database replication (DR), where replication could break if a producer cluster suffered a network partition. If the production cluster split into two segments due to network issues, a race condition could result in the consumer cluster querying the smaller segment of the cluster for topology information after the separation but before the smaller segment was shutdown by VoltDB's network partition detection. If this occurred, the consumer cluster would wait for the smaller segment and fail to poll the larger, surviving segment. This issue has been resolved.

VoltDB 8.1 introduced a performance improvement to the data loading utilities and bulkloader API. Unfortunately, this feature also introduced a potential error condition where, if the loader encounters a runtime error, such as an input value exceeding the maximum width of a VARCHAR column, rather than rolling back the transaction it could crash the database cluster. This issue has now been resolved. Because this bug can cause the database to stop, we strongly recommend all customers using 8.1 or 8.1.1 upgrade to 8.1.2 at they earliest convenience.

VoltDB 8.1 introduced an issue that could interfere with the resilience of a K-safe cluster. If a node failed while processing a multi-partition transaction, it was possible for the remaining nodes in the cluster to suffer a deadlock. When this happened, the warning "possible multipartition transaction deadlock detected" was reported and all subsequent multi-partition transactions would hang, along with certain system operations such as snapshots and command log truncation. This issue has now been resolved. Customers using V8.1 are strongly encouraged to update to V8.1.1 at their earliest convenience.

There was an issue where the voltadmin stop node command would not properly authenticate with the server when using Kerberos security. This issue has been resolved.

There was an issue where, if a database was upgraded to a new version of VoltDB using voltadmin shutdown --save, restarted, but then crashes unexpectedly (for example, using kill -9), the database could not restart a second time. This issue has been resolved.

Since VoltDB V7.7, it was possible for the response from a multi-partition read transaction to be "lost" during a node failure. This could only happen on K-safe clusters, where the node that failed was the multi-partition initiator (that is, the node responsible for coordinating multi-partition transactions), the node failed before the transaction completed, and the procedure call was invoked on a different node of the cluster. Under these conditions, the calling application might not receive a response from the invocation. Note that this issue only occurred on K-safe clusters, for read-only multi-partition queries, and for non-topology-aware clients. This issue has now been resolved.

There was a sporadic problem with authentication of the VoltDB Management Center (VMC) on slow network connections. Starting with VoltDB 7.9, the JSON interface uses stay-alive connections with a timeout period of 10 seconds. VMC sends requests every 5 seconds. However, on slow networks the VMC calls could be delayed beyond the timeout period, forcing the user to re-authenticate manually. The timeout period has been extended to alleviate the unexpected timeouts.

There was an issue with Kafka import where, if the database cluster was paused and then resumed, it was possible for certain Kafka records that were being processed when the database was paused to be lost. This issue has been resolved.

There was a longstanding and somewhat obscure bug involving partitioned views joined to a derived table from a sub SELECT statement. If a partitioned view did not include the table's partition column and was joined to another table derived from a sub selection (that is a SELECT statement in parentheses in the FROM clause of the main SELECT) the query could result in unexpected behavior at runtime, including possibly crashing the database. This issue has been resolved.

There must be more than one top-level group specified for the cluster

The same number of nodes must be included in each group

The number of partition copies (that is, K+1) must be a multiple of the number of top-level groups

"Fast" read consistency (<consistency> element)

Old, non-elastic partitioning (<cluster> elastic attribute)

@ProcInfo Java annotation for specifying procedure partitioning

Old shell commands (voltdb add, create, recover, and rejoin)

Ability to enable TLS/SSL encryption on the internal port. (See description above.)

There was an issue using the SQL Query tab of the web-based VoltDB Management Console (VMC) to insert or filter records if any text fields in the query contained multiple consecutive spaces. (For example, two or more leading spaces or multiple spaces between two words.) Some of the spaces were interpreted as the UTF-8 character for a non-breaking space (\u00A0) rather than ASCII code 32, causing incorrect data insertion or filtering. This issue has been resolved.

In the past it was possible for queries containing both a FULL or RIGHT OUTER JOIN and a GROUP BY operation on a floating point (FLOAT) column to produce incorrect results. This issue has been resolved.

Previously, certain combinations of UNION, ORDER BY, and LIMIT clauses in a single query could produce incorrect results. This issue has been resolved.

There was an issue with the Nagios script for monitoring replica clusters when using database replication (DR). The script could generate a series of false alarms if the replica database was idle for more than two minutes. The alarms incorrectly reported that replication was falling behind. This issue has been resolved.

Specify which interface to use for internal and external ports, respectively, using the --internalinterface and --externalinterface arguments when starting the database process with the voltdb start command.

For an individual port, specify the interface and port on the command line. For example voltdb start --client=32.31.30.29:21212.