23.153. Release 0.60
JDBC improvements
The Presto version of the JDBC DatabaseMetaData
interface now includes
proper implementations of getTables
, getSchemas
and getCatalogs
.
The JDBC driver is now always packaged as a standalone jar without any
dependencies. Previously, this artifact was published with the Maven
classifier standalone
. The new build does not publish this artifact
anymore.
USE CATALOG and USE SCHEMA
The Command Line Interface now supports USE CATALOG
and
USE SCHEMA
.
TPCH Connector
We have added a new connector that will generate synthetic data following the
TPC-H specification. This connector makes it easy to generate large datasets for
testing and bug reports. When generating bug reports, we encourage users to use
this catalog since it eases the process of reproducing the issue. The data is
generated dynamically for each query, so no disk space is used by this
connector. To add the tpch
catalog to your system, create the catalog
property file etc/catalog/tpch.properties
on both the coordinator and workers
with the following contents:
connector.name=tpch
Additionally, update the datasources
property in the config properties file,
etc/config.properties
, for the workers to include tpch
.
SPI changes
The Connector
interface now has explicit methods for supplying the services
expected by the query engine. Previously, this was handled by a generic
getService
method.
Note
This is a backwards incompatible change to Connector
in the SPI,
so if you have written a connector, you will need to update your code before
deploying this release.
Additionally, we have added the NodeManager
interface to the SPI to allow a
plugin to detect all nodes in the Presto cluster. This is important for some
connectors that can divide a table evenly between all nodes as long as the
connector knows how many nodes exist. To access the node manager, simply add
the following to the Plugin
class:
@Inject
public void setNodeManager(NodeManager nodeManager)
{
this.nodeManager = nodeManager;
}
Optimizations
DISTINCT LIMIT
For queries with the following form:
SELECT DISTINCT ...
FROM T
LIMIT N
We have added an optimization that stops the query as soon as N
distinct
rows are found.
Range predicates
When optimizing a join, Presto analyzes the ranges of the partitions on each side of a join and pushes these ranges to the other side. When tables have a lot of partitions, this can result in a very large filter with one expression for each partition. The optimizer now summarizes the predicate ranges to reduce the complexity of the filters.
Compound filters
Complex expressions involving AND
, OR
, or NOT
are now optimized by
the expression optimizer.
Window functions
Window functions with a PARTITION BY
clause are now distributed based on the
partition key.
Bug fixes
Scheduling
In the changes to schedule splits in batches, we introduced two bugs that resulted in an unbalanced workload across nodes which increases query latency. The first problem was not inspecting the queued split count of the nodes while scheduling the batch, and the second problem was not counting the splits awaiting creation in the task executor.
JSON conversion of complex Hive types
Presto converts complex Hive types (array, map, struct and union) into JSON. Previously, numeric keys in maps were converted to numbers, not strings, which is invalid as JSON only allows strings for object keys. This prevented the JSON Functions and Operators from working.
Hive hidden files
Presto will now ignore files in Hive that start with an underscore
_
or a dot.
. This matches the behavior of Hadoop MapReduce / Hive.Failures incorrectly reported as no data
Certain types of failures would result in the query appearing to succeed and return an incomplete result (often zero rows). There was a race condition between the error propagation and query teardown. In some cases, the query would be torn down before the exception made it to the coordinator. This was a regression introduced during the query teardown optimization work. There are now tests to catch this type of bug.
Exchange client leak
When a query finished early (e.g., limit or failure) and the exchange operator was blocked waiting for data from other nodes, the exchange was not be closed properly. This resulted in continuous failing HTTP requests which leaked resources and produced large log files.
Hash partitioning
A query with many
GROUP BY
items could fail due to an overflow in the hash function.Compiled NULL literal
In some cases queries with a select expression like
CAST(NULL AS varchar)
would fail due to a bug in the output type detection code in expression compiler.