ACID and Transactions in Hive

What is ACID and why should you use it?

ACID stands for four traits of database transactions:  Atomicity (an operation either succeeds completely or fails, it does not leave partial data), Consistency (once an application performs an operation the results of that operation are visible to it in every subsequent operation), Isolation (operations by one user do not cause unexpected side effects for other users), and Durability (once an operation is complete it will be preserved even in the face of machine or system failure).  These traits have long been expected of database systems as part of their transaction functionality.  

Up until Hive 0.13, atomicity, consistency, and durability were provided at the partition level.  Isolation could be provided by turning on one of the available locking mechanisms (ZooKeeper or in memory).  With the addition of transactions in Hive 0.13 it is now possible to provide full ACID semantics at the row level, so that one application can add rows while another reads from the same partition without interfering with each other.

Transactions with ACID semantics have been added to Hive to address the following use cases:

1. Streaming ingest of data.  Many users have tools such as Apache Flume, Apache Storm, or Apache Kafka that they use to stream data into their Hadoop cluster.  While these tools can write data at rates of hundreds or more rows per second, Hive can only add partitions every fifteen minutes to an hour.  Adding partitions more often leads quickly to an overwhelming number of partitions in the table.  These tools could stream data into existing partitions, but this would cause readers to get dirty reads (that is, they would see data written after they had started their queries) and leave many small files in their directories that would put pressure on the NameNode.  With this new functionality this use case will be supported while allowing readers to get a consistent view of the data and avoiding too many files.
2. Slow changing dimensions.  In a typical star schema data warehouse, dimensions tables change slowly over time.  For example, a retailer will open new stores, which need to be added to the stores table, or an existing store may change its square footage or some other tracked characteristic.  These changes lead to inserts of individual records or updates of records (depending on the strategy chosen).  Starting with 0.14, Hive is able to support this.
3. Data restatement.  Sometimes collected data is found to be incorrect and needs correction.  Or the first instance of the data may be an approximation (90% of servers reporting) with the full data provided later.  Or business rules may require that certain transactions be restated due to subsequent transactions (e.g., after making a purchase a customer may purchase a membership and thus be entitled to discount prices, including on the previous purchase).  Or a user may be contractually required to remove their customer’s data upon termination of their relationship.  Starting with Hive 0.14 these use cases can be supported via INSERT, UPDATE, and DELETE.

Limitations

• BEGIN, COMMIT, and ROLLBACK are not yet supported.  All language operations are auto-commit.  The plan is to support these in a future release.
• Only ORC file format is supported in this first release.  The feature has been built such that transactions can be used by any storage format that can determine how updates or deletes apply to base records (basically, that has an explicit or implicit row id), but so far the integration work has only been done for ORC.
• By default transactions are configured to be off.  See the Configuration section below for a discussion of which values need to be set to configure it.
• Tables must be bucketed to make use of these features.  Tables in the same system not using transactions and ACID do not need to be bucketed.
• At this time only snapshot level isolation is supported.  When a given query starts it will be provided with a consistent snapshot of the data.  There is no support for dirty read, read committed, repeatable read, or serializable.  With the introduction of BEGIN the intention is to support snapshot isolation for the duration of transaction rather than just a single query.  Other isolation levels may be added depending on user requests.
• The existing ZooKeeper and in-memory lock managers are not compatible with transactions.  There is no intention to address this issue.  See Basic Design below for a discussion of how locks are stored for transactions.

Streaming Ingest Interface

See the Streaming Data Ingest document for details on using streaming ingest.

Grammar Changes

INSERT...VALUES, UPDATE, and DELETE have been added to the SQL grammar, starting in Hive 0.14.  See LanguageManual DML for details.

Several new commands have been added to Hive's DDL in support of ACID and transactions, plus some existing DDL has been modified.  

A new command SHOW TRANSACTIONS has been added, see Show Transactions for details.

A new command SHOW COMPACTIONS has been added, see Show Compactions for details.

The SHOW LOCKS command has been altered to provide information about the new locks associated with transactions.  If you are using the ZooKeeper or in-memory lock managers you will notice no difference in the output of this command.  See Show Locks for details.

A new option has been added to ALTER TABLE to request a compaction of a table or partition.  In general users do not need to request compactions, as the system will detect the need for them and initiate the compaction.  However, if compaction is turned off for a table or a user wants to compact the table at a time the system would not choose to, ALTER TABLE can be used to initiate the compaction.  See Alter Table/Partition Compact for details.  This will enqueue a request for compaction and return.  To watch the progress of the compaction the user can use SHOW COMPACTIONS.

Basic Design

HDFS does not support in-place changes to files.  It also does not offer read consistency in the face of writers appending to files being read by a user.  In order to provide these features on top of HDFS we have followed the standard approach used in other data warehousing tools.  Data for the table or partition is stored in a set of base files.  New records, updates, and deletes are stored in delta files.  A new set of delta files is created for each transaction (or in the case of streaming agents such as Flume or Storm, each batch of transactions) that alters a table or partition.  At read time the reader merges the base and delta files, applying any updates and deletes as it reads.

Delta File Compaction

Occasionally these changes need to be merged into the base files.  To do this a set of threads has been added to the Hive metastore.  They determine when this compaction needs to be done, execute the compaction, and then clean up afterwards.  There are two types of compactions, minor and major.

• Minor compaction takes a set of existing delta files and rewrites them to a single delta file per bucket.
• Major compaction takes one or more delta files and the base file for the bucket and rewrites them into a new base file per bucket.

All compactions are done in the background and do not prevent concurrent reads and writes of the data.  After a compaction the system waits until all readers of the old files have finished and then removes the old files.

Base and Delta Directories

Previously all files for a partition (or a table if the table is not partitioned) lived in a single directory.  With these changes, any partitions (or tables) written with an ACID aware writer will have a directory for the base files and a directory for each set of delta files.

Lock Manager

A new lock manager has also been added to Hive, the DbLockManager.  This lock manager stores all lock information in the metastore.  In addition all transactions are stored in the metastore.  This means that transactions and locks are durable in the face of server failure.  To avoid clients dying and leaving transaction or locks dangling, a heartbeat is sent from lock holders and transaction initiators to the metastore on a regular basis.  If a heartbeat is not received in the configured amount of time, the lock or transaction will be aborted.

Configuration

A number of new configuration parameters have been added to the system to support transactions.

*hive.txn.max.open.batch controls how many transactions streaming agents such as Flume or Storm open simultaneously.  The streaming agent then writes that number of entries into a single file (per Flume agent or Storm bolt).  Thus increasing this value decreases the number of delta files created by streaming agents.  But it also increases the number of open transactions that Hive has to track at any given time, which may negatively affect read performance.

 **Worker threads spawn MapReduce jobs to do compactions.  They do not do the compactions themselves.  Increasing the number of worker threads will decrease the time it takes tables or partitions to be compacted once they are determined to need compaction.  It will also increase the background load on the Hadoop cluster as more MapReduce jobs will be running in the background.

***Decreasing this value will reduce the time it takes for compaction to be started for a table or partition that requires compaction.  However, checking if compaction is needed requires several calls to the NameNode for each table or partition that has had a transaction done on it since the last major compaction.  So decreasing this value will increase the load on the NameNode.

Configuration Values to Set for INSERT, UPDATE, DELETE

In addition to the new parameters listed above, some existing parameters need to be set to support INSERT, UPDATE, and DELETE.

Configuration Values to Set for Compaction

If the data in your system is not owned by the Hive user (i.e., the user that the Hive metastore runs as), then Hive will need permission to run as the user who owns the data in order to perform compactions.  If you have already set up HiveServer2 to impersonate users, then the only additional work to do is assure that Hive has the right to impersonate users from the host running the Hive metastore.  This is done by adding the hostname to hadoop.proxyuser.hive.hosts in Hadoop's core-site.xml file.  If you have not already done this, then you will need to configure Hive to act as a proxy user.  This requires you to set up keytabs for the user running the Hive metastore and add hadoop.proxyuser.hive.hosts and hadoop.proxyuser.hive.groups to Hadoop's core-site.xml file.  See the Hadoop documentation on secure mode for your version of Hadoop (e.g., for Hadoop 2.5.1 it is at Hadoop in Secure Mode).

Table Properties

If a table owner does not wish the system to automatically determine when to compact, then the table property NO_AUTO_COMPACTION can be set.  This will prevent all automatic compactions.  Manual compactions can still be done with Alter Table/Partition Compact statements.