12.192. Release 0.55
RC Binary 2-4x Gain in CPU Efficiency
Presto uses custom fast-path decoding logic for specific Hive fileformats. In this release we have added a fast path for for RCFile when usingthe Binary SerDe (LazyBinaryColumnarSerDe
). In ourmicro benchmarks, we see a gain between 2x and 4x in CPU efficiency comparedto the generic (slow) path. Since Hive data decoding accounts for asignificant portion of the CPU time, this shouldresult in measurable gains for most queries over RC Binary encoded data.Note that this optimization may not result in a reduction in latencyif your cluster is network or disk I/O bound.
Hash Distributed Aggregations
GROUP BY
aggregations are now distributed across a fixed number of machines.This is controlled by the property query.initial-hash-partitions
set inetc/config.properties
of the coordinator and workers. If the value islarger than the number of machines available during query scheduling, Prestowill use all available machines. The default value is 8
.
The maximum memory size of an aggregation is nowquery.initial-hash-partitions
times task.max-memory
.
Simple Distinct Aggregations
We have added support for the DISTINCT
argument qualifier for aggregationfunctions. This is currently limited to queries without a GROUP BY
clause andwhere all the aggregation functions have the same input expression. For example:
- SELECT count(DISTINCT country)
- FROM users
Support for complete DISTINCT
functionality is in our roadmap.
Range Predicate Pushdown
We’ve modified the connector API to support range predicates in addition to simple equality predicates.This lays the ground work for adding connectors to systems that support rangescans (e.g., HBase, Cassandra, JDBC, etc).
In addition to receiving range predicates, the connector can also communicateback the ranges of each partition for use in the query optimizer. This can be amajor performance gain for JOIN
queries where one side of the join hasonly a few partitions. For example:
- SELECT * FROM data_1_year JOIN data_1_week USING (ds)
If data_1_year
and data_1_week
are both partitioned on ds
, theconnector will report back that one table has partitions for 365 days and theother table has partitions for only 7 days. Then the optimizer will limitthe scan of the data_1_year
table to only the 7 days that could possiblematch. These constraints are combined with other predicates in thequery to further limit the data scanned.
Note
This is a backwards incompatible change with the previous connector SPI,so if you have written a connector, you will need to update your codebefore deploying this release.
json_array_get Function
The json_array_get()
function makes it simple to fetch a single element from ascalar json array.
Non-reserved Keywords
The keywords DATE
, TIME
, TIMESTAMP
, and INTERVAL
are no longerreserved keywords in the grammar. This means that you can access a columnnamed date
without quoting the identifier.
CLI source Option
The Presto CLI now has an option to set the query source. The sourcevalue is shown in the UI and is recorded in events. When using the CLI inshell scripts it is useful to set the —source
option to distinguish shellscripts from normal users.
SHOW SCHEMAS FROM
Although the documentation included the syntax SHOW SCHEMAS [FROM catalog]
,it was not implemented. This release now implements this statement correctly.
Hive Bucketed Table Fixes
For queries over Hive bucketed tables, Presto will attempt to limit scans tothe buckets that could possible contain rows that match the WHERE clause.Unfortunately, the algorithm we were using to select the buckets was notcorrect, and sometimes we would either select the wrong files or fail toselect any files. We have alignedthe algorithm with Hive and now the optimization works as expected.
We have also improved the algorithm for detecting tables that are not properlybucketed. It is common for tables to declare bucketing in the Hive metadata, butnot actually be bucketed in HDFS. When Presto detects this case, it fallback to a full scan of thepartition. Not only does this change make bucketing safer, but it makes it easierto migrate a table to use bucketing without rewriting all of the data.