I'm using SparkSQL to make fact table out of 5 dimensions. I'm facing
performance issue (job is taking several hours to complete), and even after
exhaustive googleing I see no solution. These are settings I have tried turing,
but no sucess.
sqlContext.sql("set spark.sql.shuffle.partitions=10"); // varied between 10 and
5000
sqlContext.sql("set spark.sql.autoBroadcastJoinThreshold=500000000"); // 500
MB, tried 1 GB
Most of RDDs are nicely parittions (500 partitions each), however largest
dimension is not partitioned at all (images) http://imgur.com/a/cUC3d. Below is
code I have used for making fact table.
resultDmn1.registerTempTable("Dmn1");
resultDmn2.registerTempTable("Dmn2");
resultDmn3.registerTempTable("Dmn3");
resultDmn4.registerTempTable("Dmn4");
resultDmn5.registerTempTable("Dmn5");
DataFrame resultFact = sqlContext.sql("SELECT DISTINCT\n" +
" 0 AS FactId,\n" +
" rs.c28 AS c28,\n" +
" dop.DmnId AS dmn_id_dim4,\n" +
" dh.DmnId AS dmn_id_dim5,\n" +
" op.DmnId AS dmn_id_dim3,\n" +
" du.DmnId AS dmn_id_dim2,\n" +
" dc.DmnId AS dmn_id_dim1\n" +
"FROM\n" +
" t10 rs\n" +
" JOIN\n" +
" t11 r ON rs.c29 = r.id\n" +
" JOIN\n" +
" Dmn4 dop ON dop.c26 = r.c25\n" +
" JOIN\n" +
" Dmn5 dh ON dh.Date = r.c27\n" +
" JOIN\n" +
" Dmn3 du ON du.c9 = r.c16\n" +
" JOIN\n" +
" t1 d ON r.c5 = d.id\n" +
" JOIN\n" +
" t2 di ON d.id = di.c5\n" +
" JOIN\n" +
" t3 s ON d.c6 = s.id\n" +
" JOIN\n" +
" t4 p ON s.c7 = p.id\n" +
" JOIN\n" +
" t5 o ON p.c8 = o.id\n" +
" JOIN\n" +
" Dmn1 op ON op.c1 = di.c1\n" +
" JOIN\n" +
" t9 ci ON ci.id = r.c24\n" +
" JOIN\n" +
" Dmn3 dc ON dc.c18 = ci.c23\n" +
"WHERE\n" +
" op.c2 = di.c2\n" +
" AND o.name = op.c30\n" +
" AND di.c3 = op.c3\n" +
" AND di.c4 = op.c4").toSchemaRDD();
resultFact.count();
resultFact.cache();
Dmn1 has 56 rows, dmn2 11, dmn3 10, dmn4 12, and dmn5 1275533 rows prior this
join. Everything is running on AWS EMR cluster, with 3 m3.2xlarge nodes in
cluster (master + 2 slaves).
Here is result of explain: http://pastebin.com/ZRUdUuYT
