Impala: a modern, open-source SQL engine for Hadoop | Hacker News Reader

Impala: a modern, open-source SQL engine for Hadoop

blog.acolyer.org

36 pointsmrry10 comments

Apache Phoenix: http://phoenix.apache.org/ is another SQL engine on top of Hbase/Hadoop. If you have an existing HBase cluster, Phoenix is a much easier way to put a SQL layer on top by simply mapping existing hbase tables to Phoenix tables. The icing is the ability to use JDBC tooling to query the system (Squirrel, DbVisualizer).

There's also Presto from fb, and Apache Drill. Phoenix seems to be on top of Hbase, which is more funkier that plain HDFS + SQL engine.

Besides, right now there's a huge need for analytical engines on top of HDFS (data lake or whatever you call it) - that's why there are all of these: (shamelessly stolen from an email from drill users):

- Apache Hive <https://hive.apache.org> (SQL-like, with interactive SQL thanks to the Stinger initiative) - Apache Drill <http://drill.apache.org> (ANSI SQL support) - Apache Spark <https://spark.apache.org> (Spark SQL <https://spark.apache.org/sql>, queries only, add data via Hive, RDD <https://spark.apache.org/docs/latest/api/scala/index.html#or... or Parquet <http://parquet.io/>) - Apache Phoenix <http://phoenix.apache.org> (built atop Apache HBase <http://hbase.apache.org>, lacks full transaction <http://en.wikipedia.org/wiki/Database_transaction> support, relational operators <http://en.wikipedia.org/wiki/Relational_operators> and some built-in functions) - Cloudera Impala <http://www.cloudera.com/content/cloudera/en/products-and-ser... (significant HiveQL support, some SQL language support, no support for indexes on its tables, importantly missing DELETE, UPDATE and INTERSECT; amongst others) - Presto <https://github.com/facebook/presto> from Facebook (can query Hive, Cassandra <http://cassandra.apache.org>, relational DBs &etc. Doesn't seem to be designed for low-latency responses across small clusters, or support UPDATE operations. It is optimized for data warehousing or analytics¹ <http://prestodb.io/docs/current/overview/use-cases.html>) - SQL-Hadoop <https://www.mapr.com/why-hadoop/sql-hadoop> via MapR community edition <https://www.mapr.com/products/hadoop-download> (seems to be a packaging of Hive, HP Vertica <http://www.vertica.com/hp-vertica-products/sqlonhadoop>, SparkSQL, Drill and a native ODBC wrapper <http://package.mapr.com/tools/MapR-ODBC/MapR_ODBC>) - Apache Kylin <http://www.kylin.io> from Ebay (provides an SQL interface and multi-dimensional analysis [OLAP <http://en.wikipedia.org/wiki/OLAP>], "… offers ANSI SQL on Hadoop and supports most ANSI SQL query functions". It depends on HDFS, MapReduce, Hive and HBase; and seems targeted at very large data-sets though maintains low query latency) - Apache Tajo <http://tajo.apache.org> (ANSI/ISO SQL standard compliance with JDBC <http://en.wikipedia.org/wiki/JDBC> driver support [benchmarks against Hive and Impala <http://blogs.gartner.com/nick-heudecker/apache-tajo-enters-t... ]) - Cascading <http://en.wikipedia.org/wiki/Cascading_%28software%29>'s Lingual <http://docs.cascading.org/lingual/1.0/>² <http://docs.cascading.org/lingual/1.0/#sql-support> ("Lingual provides JDBC Drivers, a SQL command shell, and a catalog manager for publishing files [or any resource] as schemas and tables.")

I started using Impala this summer on AWS Elastic Mapreduce. Previously I'd been using Hive. We had at least a 100x speedup in most of the work we do because of it. Things that would have been impossible before (near realtime classification) are now feasible with mini-batches. My only (major) gripe is that AWS only supports several versions back. This makes it difficult to troubleshoot problems and find documentation. That said, the benefits far far outweight the cost.

impala is good but it's no match for the best analytics databases out there

In terms of performance or features? They had a paper at CIDR this year that claims they are competitive with a well known analytics DB (presumably oracle although it wasn't named explicitly).

afaik impala doesn't have multicore joins while almost any other serious db has. so if it uses 1 core out of 10, 20 to do the join then it might be 10x, 20x slower than the best

...but throughput shouldn't be a problem if you have N queries at same time (N=nr of cores)

yeah, I guess if you want throughput then it's fine

Most of those analytics databases are pretty temperamental in cloud environments, though. I haven't used Impala, but something that can run on EMR (especially if it can get better turnaround times than Redshift) is pretty interesting.

(Vertica in AWS is a tire fire. Avoid avoid avoid.)

I have started using Impala/Parquet in combination with Solr for real-time/batch analytics. Impala is very fast - it basically bypasses the entire Hadoop stack to perform its own distributed processing and IO, and Parquet is a great data-at-rest format. I have found the best way to update data holdings is by landing data in Avro format (which is a great streaming format as it carries its schema around with it), referencing it as an external table, then scheduling an Impala query to convert the current Avro bucket into Parquet. I am also looking at using Spark Streaming for real-time map/reduce style processing to value add to the data flowing into Solr/Impala. Even with a moderate sized cluster I feel like I have a lot of processing power at my fingertips, and it all just works. Hadoop is really starting to shine.

Reply on news.ycombinator.com