Glue etl job fails to write to Redshift using dynamic frame - reason ?

Question

We are observing that writing to redshift using glue dynamic frame errors out when the input file >1GB.
**Setup :**
Redshift Cluster : 2 node DC2
**Glue job**

temp_df = glueContext.create_dynamic_frame.from_options(connection_type="s3", format="csv", connection_options={"paths": [source]}, format_options={"withHeader": True, "separator": ","}, transformation_ctx="path={}".format(source)).toDF()
    
        redshift_df = DynamicFrame.fromDF(output_df, glueContext, "redshift_df")
        datasink4 = glueContext.write_dynamic_frame.from_jdbc_conf(frame=redshift_df, catalog_connection="pilot-rs", connection_options={"preactions": "truncate table tablename;", "dbtable": "tablename", "database": "dev"}, redshift_tmp_dir='s3://bucket/path/', transformation_ctx="datasink4")

**Observation :**
Code works when the input file is under 1GB. It is able to write to redshift table.
Code fails when input file size is >1Gb and job run time is around 10 mins.

**Error:**

An error occurred while calling o260.save. Timeout waiting for connection from pool

and sometimes

“An error occurred while calling o334.pyWriteDynamicFrame. Timeout waiting for connection from pool"

Portion of glue error log

Caused by: com.amazon.ws.emr.hadoop.fs.shaded.org.apache.http.conn.ConnectionPoolTimeoutException: Timeout waiting for connection from pool
    	at com.amazon.ws.emr.hadoop.fs.shaded.org.apache.http.impl.conn.PoolingHttpClientConnectionManager.leaseConnection(PoolingHttpClientConnectionManager.java:286)
    	at com.amazon.ws.emr.hadoop.fs.shaded.org.apache.http.impl.conn.PoolingHttpClientConnectionManager$1.get(PoolingHttpClientConnectionManager.java:263)

Accepted Answer

Seems like the you ran out of HTTPConnection objects that is either trying to connect to source (s3) or connect to sink (temp location of s3). I have seen issues with EMR like this before and I did set fs.s3.maxConnections to high value to increase the connection pool size.  You can increase that as mentioned here [1]. You can set the value by below

scala: sparkcontext.hadoopConfiguration.set("spark.hadoop.fs.s3.maxConnections", 1000)
    
    python: sparkcontext._jsc.hadoopConfiguration().set('spark.hadoop.fs.s3.maxConnections', '1000')

The issue might be because of large files being fetched and written to sink and thus the HTTP Connection might take longer then normal and the pool might not just be enough for you.

[1]: https://aws.amazon.com/premiumsupport/knowledge-center/emr-timeout-connection-wait/

Glue etl job fails to write to Redshift using dynamic frame - reason ?

Contenido relevante