[jira] [Created] (PARQUET-1883) int96 support in parquet-avro

["satish (Jira)" <ji...@apache.org>]

satish created PARQUET-1883:
-------------------------------

             Summary: int96 support in parquet-avro
                 Key: PARQUET-1883
                 URL: https://issues.apache.org/jira/browse/PARQUET-1883
             Project: Parquet
          Issue Type: Bug
          Components: parquet-avro
    Affects Versions: 1.10.1
            Reporter: satish


Hi

It looks like 'timestamp' is being converted to 'int64' primitive type in parquet-avro. This is incompatible with hive2. Hive throws below error 

{code:java}
Error: java.io.IOException: org.apache.hadoop.hive.ql.metadata.HiveException: java.lang.ClassCastException: org.apache.hadoop.io.LongWritable cannot be cast to org.apache.hadoop.hive.serde2.io.TimestampWritable (state=,code=0)
{code}


What does it take to write timestamp field as 'int96'? 

Hive seems to write timestamp field as int96.  See example below

{code:java}
$ hadoop jar parquet-tools-1.9.0.jar meta hdfs://timestamp_test/000000_0
creator:     parquet-mr version 1.10.6 (build 098c6199a821edd3d6af56b962fd0f1558af849b)

file schema: hive_schema
--------------------------------------------------------------------------------
ts:          OPTIONAL INT96 R:0 D:1

row group 1: RC:4 TS:88 OFFSET:4
--------------------------------------------------------------------------------
ts:           INT96 UNCOMPRESSED DO:0 FPO:4 SZ:88/88/1.00 VC:4 ENC:BIT_PACKED,RLE,PLAIN_DICTIONARY
{code}

Writing a spark dataframe into parquet format (without using avro) is also using int96.


{code:java}
scala> testDS.printSchema()
root
 |-- ts: timestamp (nullable = true)

scala> testDS.write.mode(Overwrite).save("/tmp/x");

$ parquet-tools meta /tmp/x/part-00000-99720ebd-0aea-45ac-9b8c-0eb7ad6f4e3c-c000.gz.parquet 
file:        file:/tmp/x/part-00000-99720ebd-0aea-45ac-9b8c-0eb7ad6f4e3c-c000.gz.parquet 
creator:     parquet-mr version 1.10.1 (build a89df8f9932b6ef6633d06069e50c9b7970bebd1) 
extra:       org.apache.spark.sql.parquet.row.metadata = {"type":"struct","fields":[{"name":"ts","type":"timestamp","nullable":true,"metadata":{}}]} 

file schema: spark_schema 
--------------------------------------------------------------------------------
ts:          OPTIONAL INT96 R:0 D:1

row group 1: RC:4 TS:93 OFFSET:4 
--------------------------------------------------------------------------------
ts:           INT96 GZIP DO:0 FPO:4 SZ:130/93/0.72 VC:4 ENC:RLE,PLAIN_DICTIONARY,BIT_PACKED ST:[no stats for this column]


{code}



I saw some explanation for deprecating int96 [support here|https://issues.apache.org/jira/browse/PARQUET-1870?focusedCommentId=17127963&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#comment-17127963] from [~gszadovszky]. But given hive and serialization in other parquet modules (non-avro) support int96, I'm trying to understand the reasoning for not implementing it in parquet-avro.

A bit more context: we are trying to migrate some of our data to[ hudi format|https://hudi.apache.org/]. Hudi adds a lot of efficiency for our use cases. But, when we write data using hudi, hudi uses parquet-avro and timestamp is being converted to int64. As mentioned earlier, this breaks compatibility with hive. A lot of columns in our tables have 'timestamp' as type in hive DDL.  It is almost impossible to change DDL to long as there are large number of tables and columns. 

We are happy to contribute if there is a clear path forward to support int96 in parquet-avro. Please also let me know if you are aware of a workaround in hive that can read int64 correctly as timestamp.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)