Thank you again for coming and reviewing this PR.
So far, we discussed the followings.
1. `Why are we adding this to core? Why not just the hive module?` (@rxin)
- `sql/core` module gives more benefit than `sql/hive`.
- Apache ORC library (`no-hive` version) is a general and resonably small
library designed for non-hive apps.
2. `Can we add smaller amount of new code to use this, too?` (@kiszk)
- The previous #17980 , #17924, and #17943 are the complete examples
containing this PR.
- This PR is focusing on dependency only.
3. `Why don't we then create a separate orc module? Just copy a few of the
files over?` (@rxin)
- Apache ORC library is the same with most of other data sources(CSV, JDBC,
JSON, PARQUET, TEXT) which live inside `sql/core`
- It's better to use as a library instead of copying ORC files because
Apache ORC shaded jar has many files. We had better depend on Apache ORC
community's effort until an unavoidable reason for copying occurs.
4. `I do worry in the future whether ORC would bring in a lot more jars` (@rxin)
- The ORC core library's dependency tree is aggressively kept as small as
possible. I've gone through and excluded unnecessary jars from our
dependencies. I also kick back pull requests that add unnecessary new
dependencies. (@omalley)
5. `In the long term, Spark should move to using only the vectorized reader in
ORC's core” (@omalley)
- Of course.
I’ve been waiting for new comments and discussion since last week.
Apparently, there is no further comments except the last comment(5) from Owen
in this week.
Please give your opinion if you think we need some change on the current PR
(as-is).
FYI, there is one LGTM on the PR (as-is) and no -1 so far.
Thank you again for supporting new ORC improvement in Apache Spark.
Bests,
Dongjoon.
From: Dong Joon Hyun <dh...@hortonworks.com>
Date: Friday, August 4, 2017 at 8:05 AM
To: "dev@spark.apache.org" <dev@spark.apache.org>
Cc: Apache Spark PMC <priv...@spark.apache.org>
Subject: Use Apache ORC in Apache Spark 2.3
Hi, All.
Apache Spark always has been a fast and general engine, and
supports Apache ORC inside `sql/hive` module with Hive dependency since Spark
1.4.X (SPARK-2883).
However, there are many open issues about `Feature parity for ORC with Parquet
(SPARK-20901)` as of today.
With new Apache ORC 1.4 (released 8th May), Apache Spark is able to get the
following benefits.
- Usability:
* Users can use `ORC` data sources without hive module (-Phive) like
`Parquet` format.
- Stability & Maintanability:
* ORC 1.4 already has many fixes.
* In the future, Spark can upgrade ORC library independently from Hive
(similar to Parquet library, too)
* Eventually, reduce the dependecy on old Hive 1.2.1.
- Speed:
* Last but not least, Spark can use both Spark `ColumnarBatch` and ORC
`RowBatch` together
which means full vectorization support.
First of all, I'd love to improve Apache Spark in the following steps in the
time frame of Spark 2.3.
- SPARK-21422: Depend on Apache ORC 1.4.0
- SPARK-20682: Add a new faster ORC data source based on Apache ORC
- SPARK-20728: Make ORCFileFormat configurable between sql/hive and sql/core
- SPARK-16060: Vectorized Orc Reader
I’ve made above PRs since 9th May, the day after Apache ORC 1.4 release,
but the PRs seems to need more attention of PMC since this is an important
change.
Since the discussion on Apache Spark 2.3 cadence is already started this week,
I thought it’s a best time to ask you about this.
Could anyone of you help me to proceed ORC improvement in Apache Spark
community?
Please visit the minimal PR and JIRA issue as a starter.
* https://github.com/apache/spark/pull/18640
* https://issues.apache.org/jira/browse/SPARK-21422
Thank you in advance.
Bests,
Dongjoon Hyun.
