February 6, 2018
This document provides information about known issues and limitations to the current release of VoltDB. If you encounter any problems not listed below, please be sure to report them to firstname.lastname@example.org. Thank you.
VoltDB 8.0 is a major release incorporating features from recent point releases plus new capabilities. The major new features in V8.0 include:
More Network Security — VoltDB now provides SSL/TLS encryption as an option on all inter-node and inter-cluster communication, including internal, external, and DR ports. See the chapter on "Security" in the Using VoltDB manual for details.
User-Defined Functions — It is now possible to define and declare your own functions for use in SQL statements. User-defined functions are written in Java and declared using the CREATE FUNCTION statement. See the chapter on "Creating Custom SQL Functions" in the VoltDB Guide to Performance and Customization manual for details.
Common Table Expressions — VoltDB SQL queries can now include common table expressions, using the WITH clause. Common table expressions help organize complex SQL queries and make them easier to read. VoltDB also supports recursive common table expressions, making it possible to evaluate complex tree and graph structures within a single statement. See the description of the SELECT statement in the Using VoltDB manual for details.
Kafka Enhancementss — VoltDB now supports the latest releases of Apache Kafka, by default. The Kafka export connector continues to support all Kafka versions starting with 0.8.2. For import, the Kafka import connector and the kafkaloader command line utility now support Kafka 0.10.2 and later, up through and including the recently released version 1.0.0. For customers still using earlier versions of Kafka, Kafka 8 support is available as a configurable option for both the import connector and a legacy kafkaloader8 command line tool.
Python V3 API — VoltDB now supports the use of Python V3.x for developing client applications. The VoltDB Python client library (available from GitHub) supports both Python 2.7 and 3.x.
Most of the new features and capabilities in VoltDB V8.0 do not impact existing applications. However, there are a few changes that do require minor changes to the configuration when upgrading from earlier versions. Existing customers should take note of of the following changes:
Change to default Kafka versions for import
For the kafka import connector and the kafkaloader command line utility, the default Kafka
version has changed from 0.8.2 to 0.10.2 or later. For customers already using Kafka 0.10.2 or later, there are no
changes needed to their configuration, scripts, or applications. For customers who wish to continue using the older
Kafka version 0.8.2, they will need to add the attribute
version="8" to the import
connector configuration and/or use the command line utility kafkaloader8 instead of the default
The "elastic" attribute removed from <cluster>
An artifact of an old feature provided for backwards compatibility, the
elastic attribute of the
<cluster> element in the
configuration file was disabled and deprecated several years ago. It has now been removed. Although it is unlikely any
still exist, configuration files that do include this attribute will now fail to parse. Simply remove the attribute and
The <consistency> element removed from the configuration file
<consistency> element was recently deprecated, since "fast" read
consistency no longer provides any significant performance improvement over "safe" mode but does introduce potential
risks during failure scenarios. It has now been removed from the allowable configuration file syntax. If you included
<consistency> in your configuration file, please remove it before starting a
VoltDB 8.0 cluster.
Old commands for starting VoltDB are no longer supported
VoltDB 6.6 introduced two new integrated commands, init and start, for
starting VoltDB servers. At that time the old commands (
rejoin) were deprecated. The old commands have now been removed from
the product. If you still use the older commands, please update your scripts to use the new commands as described in the
chapter "Starting the Database" in the
Using VoltDB manual.
The process for upgrading from the recent versions of VoltDB is as follows:
Shutdown the database, creating a final snapshot (using voltadmin shutdown --save).
Upgrade the VoltDB software.
Restart the database (using voltdb start).
For DR clusters, see the section on "Upgrading VoltDB Software" in the VoltDB Administrator's Guide for special considerations related to DR upgrades. If you are upgrading from versions before V6.8, see the section on "Upgrading Older Versions of VoltDB Manually" in the same manual.
For customers upgrading from V6.x or earlier releases of VoltDB, please see the V6.0 Upgrade Notes.
For customers upgrading from V5.x or earlier releases of VoltDB, please see the V5.0 Upgrade Notes.
For customers upgrading from V4.x or earlier releases of VoltDB, please see the V4.0 Upgrade Notes.
The following updates have been made to VoltDB since the last release.
TLS/SSL encryption for intra-cluster communication
VoltDB now supports encrypting communication on the internal port, the port used for communication between nodes in the cluster, using TLS/SSL encryption. Note that encrypting the internal port automatically adds latency to any operations that require inter-node communication, such as K-safety and multi-partition procedures. The actual impact depends on the configuration and application workload. It is strongly recommended you benchmark your application before enabling internal TLS/SSL on production systems. See the chapter on "Security" in the Using VoltDB manual for details.
New behavior for placement groups
Placement groups, or rack-aware provisioning, was introduced in VoltDB 5.5. Placement groups let you specify where each node is located, so in a virtualized K-safe environment multiple copies of a partition are distributed onto distinct hardware, racks, etc. However, changes in VoltDB 7.0 to optimize K-safe partitioning in all cases ended up superceding placement groups and invalidating the rack-aware positioning.
This unintentional side effect has been corrected and placement groups once again provide rack-aware provisioning. However, the algorithm for interpreting placement groups has changed. Where before you could use a hierarchical list of names separated by periods (such as rack1.switch3.server5) the new algorithm focuses on the first name only and subnames are largely ignored.
Use of simple (non-hierarchical) placement names is recommended. In addition, the following rules apply to the top-level names:
Kafka 0.10.2 is now the default for Kafka import and kafkaloader
The default for the Kafka import connector and the kafkaloader command line utility has changed to support Kafka 0.10.2 and later, including the recent 1.0.0 release. Earlier versions of Kafka (0.8.2) are still supported through configuration options and an alternate kafkaloader8 utility.
Support for common table expressions
Deprecated features removed from the product
The following features, that had previously been deprecated, have now been removed from the product as of VoltDB 8.0:
We are also deprecating the VoltDB Deployment Manager as of VoltDB 8.0.
The following changes have been made to improve security and eliminate potential threats:
In addition to the new features and capabilities described above, the following limitations in previous versions have been resolved:
The following are known limitations to the current release of VoltDB. Workarounds are suggested where applicable. However, it is important to note that these limitations are considered temporary and are likely to be corrected in future releases of the product.
Do not use the subfolder name "segments" for the command log snapshot directory.
VoltDB reserves the subfolder "segments" under the command log directory for storing the actual command log files. Do not add, remove, or modify any files in this directory. In particular, do not set the command log snapshot directory to a subfolder "segments" of the command log directory, or else the server will hang on startup.
Some DR data may not be delivered if master database nodes fail and rejoin in rapid succession.
Because DR data is buffered on the master database and then delivered asynchronously to the replica, there is always the danger that data does not reach the replica if a master node stops. This situation is mitigated in a K-safe environment by all copies of a partition buffering on the master cluster. Then if a sending node goes down, another node on the master database can take over sending logs to the replica. However, if multiple nodes go down and rejoin in rapid succession, it is possible that some buffered DR data — from transactions when one or more nodes were down — could be lost when another node with the last copy of that buffer also goes down.
If this occurs and the replica recognizes that some binary logs are missing, DR stops and must be restarted.
To avoid this situation, especially when cycling through nodes for maintenance purposes, the key is to ensure that all buffered DR data is transmitted before stopping the next node in the cycle. You can do this using the @Statistics system procedure to make sure the last ACKed timestamp (using @Statistitcs DR on the master cluster) is later than the timestamp when the previous node completed its rejoin operation.
Avoid bulk data operations within a single transaction when using database replication
Bulk operations, such as large deletes, inserts, or updates are possible within a single stored procedure. However, if the binary logs generated for DR are larger than 45MB, the operation will fail. To avoid this situation, it is best to break up large bulk operations into multiple, smaller transactions. A general rule of thumb is to multiply the size of the table schema by the number of affected rows. For deletes and inserts, this value should be under 45MB to avoid exceeding the DR binary log size limit. For updates, this number should be under 22.5MB (because the binary log contains both the starting and ending row values for updates).
Database replication ignores resource limits
There are a number of VoltDB features that help manage the database by constraining memory size and resource utilization. These features are extremely useful in avoiding crashes as a result of unexpected or unconstrained growth. However, these features could interfere with the normal operation of DR when passing data from one cluster to another, especially if the two clusters are different sizes. Therefore, as a general rule of thumb, DR overrides these features in favor of maintaining synchronization between the two clusters.
Specifically, DR ignores any resource monitor limits defined in the deployment file when applying binary logs on the consumer cluster. DR also ignores any partition row limits defined in the database schema when applying binary logs. This means, for example, if the replica database in passive DR has less memory or fewer unique partitions than the master, it is possible that applying binary logs of transactions that succeeded on the master could cause the replica to run out of memory. Note that these resource monitor and tables row limits are applied on any original transactions local to the cluster (for example, transactions on the master database in passive DR).
Different cluster sizes can require additional Java heap
Database Replication (DR) now supports replication across clusters of different sizes. However, if the replica cluster is smaller than the master cluster, it may require a significantly larger Java heap setting. Specifically, if the replica has fewer unique partitions than the master, each partition on the replica must manage the incoming binary logs from more partitions on the master, which places additional pressure on the Java heap.
A simple rule of thumb is that the worst case scenario could require an additional P * R * 20MB space in the Java heap , where P is the number of sites per host on the replica server and R is the ratio of unique partitions on the master to partitions on the replica. For example, if the master cluster is 5 nodes with 10 sites per host and a K factor of 1 (i.e. 25 unique partitions) and the replica cluster is 3 nodes with 8 sites per host and a K factor of 1 (12 unique partitions), the Java heap on the replica cluster may require approximately 320MB of additional space in the heap:
Sites-per-host * master/replace ratio * 20MB
An alternative is to reduce the size of the DR buffers on the master cluster by setting the DR_MEM_LIMIT Java property. For example, you can reduce the DR buffer size from the default 10MB to 5MB using the VOLTDB_OPTS environment variable before starting the master cluster.
$ export VOLTDB_OPTS="-DDR_MEM_LIMIT=5"
$ voltdb start
Changing the DR buffer limit on the master from 10MB to 5MB proportionally reduces the additional heap size needed. So in the previous example, the additional heap on the replica is reduced from 320MB to 160MB.
Avoid replicating tables without a unique index.
Part of the replication process for XDCR is to verify that the record's starting and ending states match on both clusters, otherwise known as conflict resolution. To do that, XDCR must find the record first. Finding uniquely indexed records is efficient; finding non-unique records is not and can impact overall database performance.
To make you aware of possible performance impact, VoltDB issues a warning if you declare a table as a DR table and it does not have a unique index.
When starting XDCR for the first time, only one database can contain data.
You cannot start XDCR if both databases already have data in the DR tables. Only one of the two participating databases can have preexisting data when DR starts for the first time.
During the initial synchronization of existing data, the receiving database is paused.
When starting XDCR for the first time, where one database already contains data, a snapshot of that data is sent to the other database. While receiving and processing that snapshot, the receiving database is paused. That is, it is in read-only mode. Once the snapshot is completed and the two database are synchronized, the receiving database is automatically unpaused, resuming normal read/write operations.
A large number of multi-partition write transactions may interfere with the ability to restart XDCR after a cluster stops and recovers.
Normally, XDCR will automatically restart where it left off after one of the clusters stops and recovers from its command logs (using the voltdb recover command). However, if the workload is predominantly multi-partition write transactions, a failed cluster may not be able to restart XDCR after it recovers. In this case, XDCR must be restarted from scratch, using the content from one of the clusters as the source for synchronizing and recreating the other cluster (using the voltdb create --force command) without any content in the DR tables.
A TRUNCATE TABLE transaction will be reported as a conflict with any other write operation to the same table.
When using XDCR, if the binary log from one cluster includes a TRUNCATE TABLE statement and the other cluster performs any write operation to the same table before the binary log is processed, the TRUNCATE TABLE operation will be reported as a conflict. Note that currently DELETE operations always supercede other actions, so the TRUNCATE TABLE will be executed on both clusters.
Exceeding a LIMIT PARTITION ROWS constraint can generate multiple conflicts
It is possible to place a limit on the number of rows that any partition can hold for a specific table using the LIMIT PARTITION ROWS clause of the CREATE TABLE statement. When close to the limit, transactions on either or both clusters can exceed the limit simultaneously, resulting in a potentially large number of delete operations that then generate conflicts when the the associated binary log reaches the other cluster.
Use of the VoltProcedure.getUniqueId method is unique to a cluster, not across clusters.
VoltDB provides a way to generate a deterministically unique ID within a stored procedure using the getUniqueId method. This method guarantees uniqueness within the current cluster. However, the method could generate the same ID on two distinct database clusters. Consequently, when using XDCR, you should combine the return values of VoltProcedure.getUniqueId with VoltProcedure.getClusterId, which returns the current cluster's unique DR ID, to generate IDs that are unique across all clusters in your environment.
XDCR cannot be used with deprecated export syntax.
You cannot use cross datacenter replication (XDCR) with the deprecated export syntax, that is the EXPORT TABLE statement. To use XDCR with export, you must use the current CREATE STREAM syntax for declaring the source streams for export targets.
Synchronous export in Kafka can use up all available file descriptors and crash the database.
A bug in the Apache Kafka client can result in file descriptors being allocated but not released if the producer.type attribute is set to "sync" (which is the default). The consequence is that the system eventually runs out of file descriptors and the VoltDB server process will crash.
Until this bug is fixed, use of synchronous Kafka export is not recommended. The workaround is to set the Kafka producer.type attribute to "async" using the VoltDB export properties.
Data may be lost if a Kafka broker stops during import.
If, while Kafka import is enabled, the Kafka broker that VoltDB is connected to stops (for example, if the server crashes or is taken down for maintenance), some messages may be lost between Kafka and VoltDB. To ensure no data is lost, we recommend you disable VoltDB import before taking down the associated Kafka broker. You can then re-enable import after the Kafka broker comes back online.
Kafka import can lose data if multiple nodes stop in succession.
There is an issue with the Kafka importer where, if multiple nodes in the cluster fail and restart, the importer can lose track of some of the data that was being processed when the nodes failed. Normally, these pending imports are replayed properly on restart. But if multiple nodes fail, it is possible for some in-flight imports to get lost. This issue will be addressed in an upcoming release.
Comments containing unmatched single quotes in multi-line statements can produce unexpected results.
When entering a multi-line statement at the sqlcmd prompt, if a line ends in a comment (indicated by two hyphens) and the comment contains an unmatched single quote character, the following lines of input are not interpreted correctly. Specifically, the comment is incorrectly interpreted as continuing until the next single quote character or a closing semi-colon is read. This is most likely to happen when reading in a schema file containing comments. This issue is specific to the sqlcmd utility.
A fix for this condition is planned for an upcoming point release
Do not use assertions in VoltDB stored procedures.
VoltDB currently intercepts assertions as part of its handling of stored procedures. Attempts to use assertions in stored procedures for debugging or to find programmatic errors will not work as expected.
The UPPER() and LOWER() functions currently convert ASCII characters only.
The UPPER() and LOWER() functions return a string converted to all uppercase or all lowercase letters, respectively. However, for the initial release, these functions only operate on characters in the ASCII character set. Other case-sensitive UTF-8 characters in the string are returned unchanged. Support for all case-sensitive UTF-8 characters will be included in a future release.
Avoid using decimal datatypes with the C++ client interface on 32-bit platforms.
There is a problem with how the math library used to build the C++ client library handles large decimal values on 32-bit operating systems. As a result, the C++ library cannot serialize and pass Decimal datatypes reliably on these systems.
Note that the C++ client interface can send and receive Decimal values properly on 64-bit platforms.
Enabling SNMP traps can slow down database startup.
Enabling SNMP can take up to 2 minutes to complete. This delay does not always occur and can vary in length. If SNMP is enabled when the database server starts, the delay occurs after the server logs the message "Initializing SNMP" and before it attempts to connect to the cluster. If you enable SNMP while the database is running, the delay can occur when you issue the voltadmin update command or modify the setting in the VoltDB Management Center Admin tab. This issue results from a Java constraint related to secure random numbers used by the SNMP library.
The VoltDB Management Center currently reports on only one DR connection.
With VoltDB V7.0, cross datacenter replication (XDCR) supports multiple clusters in an XDCR network. However, the VoltDB Management Center currently reports on only one such connection per cluster. In the future, the Management Center will provide monitoring and statistics for all connections to the current cluster.
The following notes provide details concerning how certain VoltDB features operate. The behavior is not considered incorrect. However, this information can be important when using specific components of the VoltDB product.
Schema updates clear the stored procedure data table in the Management Center Monitor section
Any time the database schema or stored procedures are changed, the data table showing stored procedure statistics at the bottom of the Monitor section of the VoltDB Management Center get reset. As soon as new invocations of the stored procedures occur, the statistics table will show new values based on performance after the schema update. Until invocations occur, the procedure table is blank.
You cannot partition a table on a column defined as ASSUMEUNIQUE.
The ASSUMEUNIQUE attribute is designed for identifying columns in partitioned tables where the column values are known to be unique but the table is not partitioned on that column, so VoltDB cannot verify complete uniqueness across the database. Using interactive DDL, you can create a table with a column marked as ASSUMEUNIQUE, but if you try to partition the table on the ASSUMEUNIQUE column, you receive an error. The solution is to drop and add the column using the UNIQUE attribute instead of ASSUMEUNIQUE.
Adding or dropping column constraints (UNIQUE or ASSUMEUNIQUE) is not supported by the ALTER TABLE ALTER COLUMN statement.
You cannot add or remove a column constraint such as UNIQUE or ASSUMEUNIQUE using the ALTER TABLE ALTER COLUMN statement. Instead to add or remove such constraints, you must first drop then add the modified column. For example:
ALTER TABLE employee DROP COLUMN empID; ALTER TABLE employee ADD COLUMN empID INTEGER UNIQUE;
Do not use UPDATE to change the value of a partitioning column
For partitioned tables, the value of the column used to partition the table determines what partition the row belongs to. If you use UPDATE to change this value and the new value belongs in a different partition, the UPDATE request will fail and the stored procedure will be rolled back.
Updating the partition column value may or may not cause the record to be repartitioned (depending on the old and new values). However, since you cannot determine if the update will succeed or fail, you should not use UPDATE to change the value of partitioning columns.
The workaround, if you must change the value of the partitioning column, is to use both a DELETE and an INSERT statement to explicitly remove and then re-insert the desired rows.
Certain SQL syntax errors result in the error message "user lacks privilege or object not found".
If you refer to a table or column name that does not exist, VoltDB reports that the "user lacks privilege or object not found". This can happen, for example, if you misspell a table or column name.
Another situation where this occurs is if you mistakenly use double quotation marks to enclose a string
literal (such as
The workaround is, if you receive this error, to look for misspelled table or columns names or string literals delimited by double quotes in the offending SQL statement.
Ambiguous column references no longer allowed.
Starting with VoltDB 6.0, ambiguous column references are no longer allowed. For example, if both the Customer and Placedorder tables have a column named Address, the reference to Address in the following SELECT statement is ambiguous:
SELECT OrderNumber, Address FROM Customer, Placedorder . . .
Previously, VoltDB would select the column from the leftmost table (Customer, in this case). Ambiguous column references are no longer allowed and you must use table prefixes to disambiguate identical column names. For example, specifying the column in the preceding statement as Customer.Address.
A corollary to this change is that a column declared in a USING clause can now be referenced using a prefix. For example, the following statement uses the prefix Customer.Address to disambiguate the column selection from a possibly similarly named column belonging to the Supplier table:
SELECT OrderNumber, Vendor, Customer.Address FROM Customer, Placedorder Using (Address), Supplier . . .
File Descriptor Limits
VoltDB opens a file descriptor for every client connection to the database. In normal operation, this use of file descriptors is transparent to the user. However, if there are an inordinate number of concurrent client connections, or clients open and close many connections in rapid succession, it is possible for VoltDB to exceed the process limit on file descriptors. When this happens, new connections may be rejected or other disk-based activities (such as snapshotting) may be disrupted.
In environments where there are likely to be an extremely large number of connections, you should consider increasing the operating system's per-process limit on file descriptors.
Use of Resources in JAR Files
There are two ways to access additional resources in a VoltDB database. You can place the resources in the
LOAD CLASSES is used primarily to load classes associated with stored procedures and user-defined functions. However, it will also load any additional resource files included in subfolders of the JAR file. You can remove classes that are no longer needed using the REMOVE CLASSES directive. However, there is no explicit command for removing other resources.
Consequently, if you rename resources or move them to a different location and reload the JAR file, the database will end up having multiple copies. Over time, this could result in more and more unnecessary memory being used by the database. To remove obsolete resources, you must first reinitialize the database root directory, start a fresh database, reload the schema (including the new JAR files with only the needed resources) and then restore the data from a snapshot.
Servers with Multiple Network Interfaces
If a server has multiple network interfaces (and therefore multiple IP addresses) VoltDB will, by default, open ports on all available interfaces. You can limit the ports to an single interface in two ways:
Also, when using an IP address to reference a server with multiple interfaces in command line utilities (such as voltadmin stop node), use the @SystemInformation system procedure to determine which IP address VoltDB has selected to identify the server. Otherwise, if you choose the wrong IP address, the command might fail.