mbutrovich commented on code in PR #17162:
URL: https://github.com/apache/iceberg/pull/17162#discussion_r3570946768


##########
site/docs/blog/posts/2026-07-10-accelerating-iceberg-rust-development-with-datafusion-comet.md:
##########
@@ -0,0 +1,213 @@
+---
+date: 2026-07-10
+title: Accelerating Apache Spark Queries (and Iceberg Rust Development) with 
Apache DataFusion Comet
+slug: accelerating-iceberg-rust-development-with-datafusion-comet  # this is 
the blog url
+authors:
+  - mbutrovich
+categories:
+  - blog
+---
+
+<!--
+ - Licensed to the Apache Software Foundation (ASF) under one or more
+ - contributor license agreements.  See the NOTICE file distributed with
+ - this work for additional information regarding copyright ownership.
+ - The ASF licenses this file to You under the Apache License, Version 2.0
+ - (the "License"); you may not use this file except in compliance with
+ - the License.  You may obtain a copy of the License at
+ -
+ -   http://www.apache.org/licenses/LICENSE-2.0
+ -
+ - Unless required by applicable law or agreed to in writing, software
+ - distributed under the License is distributed on an "AS IS" BASIS,
+ - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ - See the License for the specific language governing permissions and
+ - limitations under the License.
+ -->
+
+<!-- more -->
+
+Apache Iceberg provides a universal table format that serves as a foundation 
for modern data
+lakehouse
+platforms. With Iceberg, users store their tables with the benefit of being 
able to access
+and modify their data from a number of different query engines.
+[Apache Spark](https://spark.apache.org) is the engine most closely associated 
with Iceberg. The
+[Iceberg Java repository](https://github.com/apache/iceberg), the *de facto* 
reference
+implementation of the Iceberg specification, ships Spark as its most mature 
integration. It is also the
+engine most teams rely on for table maintenance like compaction and snapshot 
expiration.
+In addition to Java, the Iceberg community maintains a number
+of other Iceberg implementations like 
[C++](https://github.com/apache/iceberg-cpp),
+[Go](https://github.com/apache/iceberg-go), and 
[Rust](https://github.com/apache/iceberg-rust).
+These other implementations benefit not only from the Iceberg specification, 
but also the lessons
+learned and design decisions of the Java project's community. The Java 
repository's extensive
+test suites, for instance, include nearly 10,000 correctness tests driven by 
Spark (as of Iceberg
+1.11 with Spark 4.1). Each implementation maintains its own test suite and can 
look to Iceberg Java
+as a reference for both correct behavior and test coverage. None of them, 
however, can run Java's
+tests directly against their own code.
+
+While Spark remains a powerful and robust engine, a number of projects exist 
to accelerate its
+JVM-backed execution. One such solution is
+[Apache DataFusion Comet](https://datafusion.apache.org/comet/), which Apple 
donated in 2024
+as a subproject of the [Apache DataFusion](https://datafusion.apache.org) 
query engine. Comet's
+native execution engine runs CPU-bound jobs faster and IO-bound jobs with
+fewer resources, giving users control over how they want to optimize their 
Spark jobs. As we will
+see, Comet does more than speed up queries: the same design that makes it fast 
also makes it a tool
+for accelerating Iceberg Rust's development.
+
+## Accelerating Spark Queries with Comet
+
+Comet builds upon several related Apache projects including DataFusion (for 
its efficient operator
+implementations like joins and aggregations), 
[Arrow-rs](https://github.com/apache/arrow-rs)
+(for its standardized in-memory format and robust Parquet reader), and, 
somewhat surprisingly,
+both the Java and the Rust implementations of Iceberg. To accelerate Spark 
queries, Comet
+intercepts execution
+at the physical plan level. After Spark has parsed, planned, and optimized a 
user's query,
+Comet's JVM code runs as one final optimizer rule to convert Spark plan nodes 
to Comet plan nodes.
+These Comet plan nodes have a superpower: they execute in DataFusion's Rust 
engine over columnar Arrow
+data.
+
+<figure markdown="span">![Comet converts a Spark physical plan into an 
equivalent DataFusion physical 
plan](../../assets/images/2026-07-10-comet-plan-translation.png){ width="750" 
}<figcaption>Comet converts a Spark physical plan into an equivalent DataFusion 
physical plan.</figcaption></figure>
+
+So how does Comet use *both* Iceberg libraries to accelerate Spark queries 
over Iceberg tables?
+As previously mentioned, Iceberg provides robust integrations with Spark, 
enabling users to query
+their Iceberg tables regardless of the Spark API they are using (*e.g.*, SQL, 
Scala, or PySpark).
+Iceberg relies on Spark's
+[`Data Source 
v2`](https://spark.apache.org/docs/4.2.0-preview4/sql-data-sources-v2.html) API 
to
+integrate with query planning, a process that
+Apache Iceberg PMC member Russell Spitzer recently described in a talk titled 
+["An Extremely Technical Overview of How Apache Iceberg Planning Actually 
Works"](https://www.youtube.com/watch?v=kJaD0WuQ1Bg).
+The short version of the talk is that given a query reading an Iceberg table, 
the Java 
+library inspects table metadata (*e.g.*, version history, schema, statistics, 
file layout) to
+construct `FileScanTask` objects. These objects describe the low-level 
operations (*e.g.*, file paths
+and byte ranges) needed to read the table and feed data to downstream query 
operators.
+
+Comet still relies on Iceberg Java for this planning. Acceleration is possible 
because Iceberg Rust
+has its own `FileScanTask`, so Comet uses it as the common abstraction between 
the two libraries: it
+takes the `FileScanTask` objects that Iceberg Java produced and hands them to 
Iceberg Rust, which
+reads the described files into the in-memory Arrow batches that feed the rest 
of the plan.
+
+<figure markdown="span">![Comet translates Iceberg Java's FileScanTask objects 
into Iceberg Rust's FileScanTask 
objects](../../assets/images/2026-07-10-comet-task-translation.png){ 
width="750" }<figcaption>Comet translates Iceberg Java's 
<code>FileScanTask</code> objects into Iceberg Rust's <code>FileScanTask</code> 
objects.</figcaption></figure>
+
+To measure Comet's impact on real workloads, the AWS Data on EKS team 
benchmarked Comet against
+Spark alone on the TPC-DS 3 TB workload over Iceberg tables. Comet completed 
the suite roughly
+40% faster (2,803.80s versus 4,665.47s) and accelerated 102 of the 103 TPC-DS 
queries, with only
+a single query regressing. See the
+[full benchmark 
writeup](https://awslabs.github.io/data-on-eks/docs/benchmarks/spark-datafusion-comet-benchmark)
+for the complete methodology and per-query results.
+
+<figure markdown="span">![TPC-DS 3 TB on Iceberg: Spark with Comet completes 
in 2803.8s versus 4665.47s for Spark with 
Iceberg](../../assets/images/2026-07-10-comet-tpcds.png){ width="500" 
}<figcaption>TPC-DS 3 TB (Iceberg) on AWS EKS: Spark with Comet completes the 
suite ~40% faster.</figcaption></figure> 
+
+Raw speed only matters if the answers are correct. Comet prioritizes 
correctness and compatibility
+with the libraries it accelerates. In addition to its own exhaustive test 
suites, Comet goes
+further by running Iceberg Java's Spark test suites with Comet enabled as 
regression tests,
+continuously checking the native path against the same corpus that guards the 
reference
+implementation.
+
+Comet does not yet accelerate all Iceberg table reads. For example, Comet 
currently falls back to

Review Comment:
   Comet doesn't accelerate writes at all right now. The iceberg-rust write 
path is even more immature. Maybe I can raise that as future green pastures to 
use this development approach to bootstrap as well (phrased better).



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to