I’ve been working on MapReduce frameworks since mid 2005 (Hadoop’s since the start of 2006) and a fundamental feature has always been incredible throughput to access data, but no ACID transactions. That is changing.
Recently, while working with a customer that is using Apache Hive to process terabytes (and growing quickly) of sales data, they asked how to handle a business requirement to update millions of records in their sales table each day. Hive doesn’t provide the right features to satisfy that requirement, however it is extensible, which enabled the customer create a custom InputFormat and OutputFormat that did. So, rather than building a new SQL engine on Hadoop, they solved the problem by customizing Hive.
We want to provide that generalized UPDATE capability to the rest of the Hive community without creating a new SQL engine on top of Hadoop.
The material below represents a general approach for incorporating ACID transactions in Hive. It represents our thoughts with the feedback of the broad Hive community. A jira ticket has been entered (HIVE-5317), that outlines the requirements and general approach and we are excited that the broad community has taken part in the discussion of how to complete this important project.
It’s clear that there is a desire for update or delete records at less than partition granularity. Furthermore, although Hive can insert to a partition, such additions are visible immediately when written instead of when the transaction completes. Typical relevant use cases are:
It is critical that we integrate this functionality into Hive using standard SQL syntax. In particular, over the course of the project, we’ll add the following standard SQL statements to Hive:
INSERT INTO tbl SELECT …
INSERT INTO tbl VALUES …
UPDATE tbl SET … WHERE …
DELETE FROM tbl WHERE …
MERGE INTO tbl USING src ON … WHEN MATCHED THEN … WHEN NOT MATCHED THEN …
SET TRANSACTION LEVEL …
It is easy to overlook the fact that adding insert, update, and delete commands will exacerbate Hive’s lack of transaction model. Therefore, a critical part of the work will be adding a transaction model to Hive. Hive’s metastore will track the read and write transactions and assign a globally unique transaction id to each write transaction. Read operations will be given the list of completed transaction ids and will ignore the output of all other transactions, providing a consistent view of all of the tables, even if some of their tasks are re-executed after a node failure.
When a read operation starts, it provides the tables it is reading and requests the current set of committed transaction ids. This set will be represented as the maximum committed transaction
id and the list of transaction ids that are still in flight. Each MapReduce job launched as part of the query will have exactly the same set of valid transaction ids provided in the JobConf and thus the reader will present a consistent snapshot view of the input tables from the time at the start of the command.
Write operations will provide the list of tables being read and the list of tables being written and get a list of valid transaction ids to read from and a write transaction id. Any data that is read will only include the valid transaction ids or their write transaction id. All data that is written will be tagged with the write transaction id.
In both cases, when the command is finished, it must notify the metastore. Since the metastore must identify (and abort) abandoned transactions, the Hive client must notify the metastore that the command is still running every 10 minutes or the transaction will be aborted. The metastore will maintain the list of tables being read and the corresponding transaction ids so that it can schedule compacted data to be deleted after the last reader finishes.
Before a write transaction is committed in the metastore, all of the HDFS files for that transaction must be in their final location. Similarly, before a write transaction is aborted in the metastore all of the files must have been deleted from HDFS. This ensures that any valid transaction id that is given to a reader will only have correct HDFS files associated with it.
The heart of the approach is merging the HDFS in the tasks. Hive’s warehouse layout will remain the same and look like $database/$table/$partition. Inside of each partition (or the table directory, if the table isn’t partitioned), instead of a set of files named with the bucket id (eg. 000000_0) there will be a directories with the base rows and the deltas that have modified that base. The base files will be stored in a directory named base_$etid and the deltas will be stored in directories named delta_$btid_$etid, where $btid is the first transaction id included in the file and $etid is the first transaction id not included in the file. This naming scheme will enable us to leave both the pre and post compaction files while the clients using the pre compaction files finish. Within each directory, the rows must be consistently bucketed so that each bucket can be processed independently.
In theory the base files can be in any format, but ORC will be required for the first version. Every row in the table is uniquely identified by the transaction id of the transaction that inserted it (or 0 for the base file), the row id, and the implicit bucket number. The implementation of delta files will be ORC files with a row format of:
create table deltaFile (
operation int, — 0 = insert, 1 = delete, 2 = update
transactionId long, — transaction that inserted the row
rowId long, — row id
rowData struct(…) — for insert and update, the columns
For sorted tables, the base and delta files will be sorted by the sort columns, while unsorted tables will be sorted by descending transaction id and ascending row id. Since the base and delta files are sorted identically, it is a simple merge sort to combine them. ORC’s integer run length encoding will compress the delta files very effectively.
When the read query is done, it will inform the metastore that it is no longer reading the tables so that obsolete versions of the tables can be garbage collected.
When one of the insert, update, or delete commands starts, it gets the valid transaction ids for reading and a transaction id to write to. The first version of the project will do auto-commit of the transaction when each query finishes. Because each of the delta files is labeled with the transaction ids, none of the data associated with a transaction is read before the transaction is committed.
The delta files are written as ORC files with a smaller stripe size (32MB), a smaller buffer size (32K), and no compression. This is because many of the files will be very small and lowering the sizes will reduce the resource requirements for reading the files.
For the insert command, the tasks will write the delta file with the insert rows. Meanwhile, updates and deletes will read the current table with the record reader and will write a new delta file with the delete and update commands.
While any query is running, the client will heartbeat to the metastore every N minutes to ensure the transaction isn’t removed for inactivity. After the command finishes, the transaction will be committed by the Hive metastore or be aborted.
Periodically the system must rewrite the smaller delta files into larger ones (minor compaction) and rewriting the delta files into the base file (major compaction). Since the compactions may take relatively large amount of time, they should not block other updates. This is possible because although they are rewriting the data, they are semantically equivalent of the previous data.
To preserve read throughput over time, compactions must happen regularly and therefore automatically. In particular, since reading the table involves doing a N-way merge sort, we must bound the size of N to a relatively small number. Minor compactions should be automatically scheduled for a table when we have more than 10 committed deltas. The minor compaction
asks the metastore for the smallest inflight transaction id and will compact all transactions that are strictly less than it. Once the minor compaction is done, the directory with the output is moved into the partition’s directory with the name delta_$bid_$eid. The previous delta directories will be scheduled for garbage collection once the transactions using the previous version have finished.
When more than 10% of the records are coming from delta files instead of the base, a major compaction should be triggered. We will also need a time based trigger to do major compactions during times when the cluster is less busy. The major compaction will request the lowest inflight transaction id and rewrite the base with the merged transaction that are less than it. When the compaction is done, the output is moved to base_$etid and the old base is scheduled for garbage collection. Because compaction will change the row ids, the major compaction will also output a set of files that translate from the previous major compaction to the new one. The translation files will be stored in the base directory and be named base_$etid/_translation_$bucketid. To keep the translation files small, it is sufficient to write the old transaction id, row id, and the number of sequential rows. With this encoding, the translation files should be proportional to the size of the deltas rather than the size of the base. After the major compaction is finished, if there are deltas using the old base file the system should schedule a minor compaction to be run. When the minor compaction is done, the translation files can be scheduled for garbage collection. Until the minor compaction finishes, the record reader will need to merge the translation files together with the old delta files to find the updated or deleted records.
One of the questions that comes up when we discuss this project is why don’t we just use Hive over HBase, since HBase already supports updates to values. There are several different issues that we’ve observed:
Since we’ve filed the Jira on Hive Updates there has been a great discussion on the proposal from the community at large and drawn an audience of 41 watchers. The approach is validated by the active use at the original customer. We are in the process of pushing the design to the next level and starting development.
As always, we do all of our development in the open at Apache to gain strength from discussions with the community rather than dump a huge patch on the open source project just before the project’s next release.