Critical Description
When reading a field of a nested column from Parquet, SparkSQL reads and assemble all the fields of that nested column. This is unnecessary, as Parquet supports fine-grained field reads out of a nested column. This may degrades the performance significantly when a nested column has many fields.
For example, I loaded json tweets data into SparkSQL and ran the following query:
SELECT User.contributors_enabled from Tweets;
User is a nested structure that has 38 primitive fields (for Tweets schema, see: https://dev.twitter.com/overview/api/tweets), here is the log message:
14/11/19 16:36:49 INFO InternalParquetRecordReader: Assembled and processed 385779 records from 38 columns in 3976 ms: 97.02691 rec/ms, 3687.0227 cell/ms
For comparison, I also ran:
SELECT User FROM Tweets;
And here is the log message:
14/11/19 16:45:40 INFO InternalParquetRecordReader: Assembled and processed 385779 records from 38 columns in 9461 ms: 40.77571 rec/ms, 1549.477 cell/ms
So both queries load 38 columns from Parquet, while the first query only needs 1 column. I also measured the bytes read within Parquet. In these two cases, the same number of bytes (99365194 bytes) were read.
Attachments
Issue Links
- is duplicated by
-
SPARK-5446 Parquet column pruning should work for Map and Struct
-
- Resolved
-
- is related to
-
SPARK-25879 Schema pruning fails when a nested field and top level field are selected
-
- Resolved
-
- relates to
-
-
- Resolved
-
- links to
