Everett Roeth (@everettroeth) profile picture

Building stuff @LakeSail

California, USA
Joined July 2020
Photo by Everett Roeth (@everettroeth)Photo by Everett Roeth (@everettroeth)
Everett Roeth retweeted
in a world where data interoperability is a must, it's absurd to see the games dbx plays...
When someone tells you "we use Unity Catalog," what do you think they mean? Turns out there are two! • Unity Catalog OSS, an LF AI & Data sandbox-stage project • Unity Catalog, Databricks' proprietary, closed-source governance service I decided to dig into the data.
22
3
4
101
45,921
Everett Roeth retweeted
When someone tells you "we use Unity Catalog," what do you think they mean? Turns out there are two! • Unity Catalog OSS, an LF AI & Data sandbox-stage project • Unity Catalog, Databricks' proprietary, closed-source governance service I decided to dig into the data.
2
6
3
62
48,570
Everett Roeth retweeted
Why is Apache Arrow becoming the universal columnar foundation for modern data and AI infrastructure? Apache Arrow is great because of its language-independent columnar memory format, zero-copy data interchange, efficient in-memory processing, and broad multi-language ecosystem. That’s why projects like Apache Spark, Apache Parquet, ClickHouse, Dremio, Polars, Pandas, InfluxData, OpenObserve, Greptime, LakeSail, Spice AI, IOMETE, Daft, Apache Doris, and so many other great systems use it. For data and AI systems, Arrow is not just another format. It is becoming a foundational layer for moving and processing data efficiently across the entire ecosystem. RisingWave, a unified data platform for real-time and Iceberg-based lakehouse workloads, also uses Arrow extensively, particularly for interoperability with Apache Iceberg, Apache DataFusion, Parquet, UDFs, and some connectors. Shoutout to our partner @columnar_tech for advancing Apache Arrow and ADBC. They're really doing great work and redefining database access and connectivity. 🙌
1
3
1
6
602
Everett Roeth retweeted
A recent Databricks customer's workload ran 381x faster: 38 minutes to 6 seconds. Test yours in minutes.

4
61
2,143,267
Everett Roeth retweeted
A data platform that's simple, transparently priced, and ridiculously fast. The LakeSail Platform runs your Spark SQL and Python jobs on a Rust-native runtime. It's faster than Spark, drastically cuts hardware costs, and scales from a single node to a multi-node cluster. It reads and writes your Delta Lake and Iceberg tables where they already live, so your lakehouse stays open and your data stays in place. For those who want to run Sail in production most efficiently, with all the operations handled for you, this platform was built for you.
1
2
1
6
323
Everett Roeth retweeted
We just published a derived TPC-DS for Sail at 3 TB and 1 TB. Here is a comparison of our benchmark to another independently published benchmark on Spark and Spark + Gluten/Velox by AWS.
1
2
6
594
Everett Roeth retweeted
Sail ran derived TPC-DS at 3TB and 1TB across a distributed cluster, and the results are the latest evidence for our vision in rethinking distributed compute. Hadoop pioneered open-source distributed processing, built around disk. Spark moved intermediate data into memory, though its shuffle stayed disk-first. When we started Sail, we envisioned another shift in distributed compute, toward doing even more of the work in memory, with an engine built entirely in Rust. We built the first version of Sail's distributed architecture in November 2024, with Arrow Flight carrying shuffle data between workers, and redesigned the cluster control plane earlier this year. Arrow Flight shuffle can look surprising at first. A TPC-DS query with dozens of stages and a few hundred partitions can have more than a million logical shuffle streams active in the cluster. Arrow Flight is built on gRPC and HTTP/2, so one TCP connection per pair of workers carries them all. The same approach has since come up in the DataFusion Comet and Riffle communities as well, and we are glad to have started proving this direction two years ago. The full benchmark ran in 1,289 seconds at 3TB and 444 seconds at 1TB, with no shuffle data written to disk. On similar hardware and settings, Sail outperforms the published TPC-DS results for Spark and Spark accelerators at these scales. The post covers the results, the design, and the full setup: lakesail.com/blog/tpcds-benc…
1
2
4
140
Everett Roeth retweeted
Sail is the first open-source engine to ship with built-in Jev support! Check out the docs to get started: docs.lakesail.com/sail/lates…
3
8
11,157
Everett Roeth retweeted
Sail now includes Jev built natively into the engine, the first of its kind. Any Spark SQL or DataFrame query can ask Jev about every row of a table, with nothing extra to install. The answers come back as typed columns you can filter, join, and aggregate like any other data. The same query can run on your laptop or across a cluster. Check out the full release and some examples of it in action. Link in thread.
1
3
7
263
Everett Roeth retweeted
Sail 0.7.1 adds the first ever Rust implementation of Apache Celeborn as a shuffle backend. If you already run Celeborn for your Spark workloads, Sail uses that cluster as it stands. The same masters, same workers, nothing is changed on the Celeborn side. Read more on the engineering in our blog: lakesail.com/blog/sail-celeb…
2
4
155
All ahoy— @LakeSail team is coming to Amsterdam! Speaking at @pydataamsterdam and our own meetup with @adyen on Wednesday: bay.news/amsterdam1 — join us for food, drink, Rust, AI and Data!
1
6
271
Everett Roeth retweeted
The interesting part of rebuilding Spark in Rust wasn’t the Rust. @shehab_amins joined Daniel Beach on Data Engineering Central to talk about why Spark compatibility turned out to be much harder, and more interesting, than expected. And what changes when the thing querying your lakehouse is an agent instead of an analyst. Delta vs. Iceberg, streaming vs. batch, Arrow and DataFusion. Full episode below. dataengineeringcentral.subst…
3
4
6
707
Everett Roeth retweeted
Your query plan, before and after checkpoint.
1
2
4
358
Everett Roeth retweeted
Sail 0.7 has been released, with two highly requested features: blocking shuffle and checkpoint. In Spark and most engines, shuffle and checkpoint data sit on the workers themselves, so a lost worker loses data and a running job pins its cluster. Sail 0.7 moves that state out to object storage: failed tasks retry by reading persisted data back, finished stages hand their workers to the next one, and iterative workloads checkpoint instead of recomputing every loop. Check out the full release: lakesail.com/blog/sail-0-7-b…
5
2
2
8
727
If you’re curious and love to learn, there has never been a better time to be alive.
706
2,179
412
15,105
8,031,211
Everett Roeth retweeted
Sail in Spark docs would have been cool 😞
5
4
9
9,006
Everett Roeth retweeted
is it over for the JVM? 👉 join me as I talk to "Jordan has no life", a cracked distributed systems youtuber with >100k subs, about all things data infra & agent coding: • columnar processing and how the industry is standardizing on it (Apache Arrow) 🏹 • the proliferation of new file formats (Vortex, Lance, Nimble, F3) 🎁 • object storage as the backbone for modern data systems 💪 • the open core business model (e.g the Tabular acquisition and Iceberg's PMC dominance) 👀 We also touched on all the second-order effects from AI coding revolution: • English becoming the working abstraction for code • human reviewers becoming the bottleneck (especially in OSS) • focus on benchmarks, tests and design more than implementation • testing practices for the AI era - triaging flaky tests with AI, fuzz/chaos testing a-la TigerBeetle Last but not least, we touched on Jordan's two latest projects he's been working on: StreamFusion and IceStream 🌟 • StreamFusion accelerates Apache Flink by offloading Flink SQL operators to Apache DataFusion using Rust and Apache Arrow 🐿️ • IceStream, which optimizes streaming Iceberg upserts through an Apache Paimon-backed index to asynchronously convert Iceberg equality deletes 🧊 TIMELINE 0:00 the future of the Data Ecosystem: new file formats, columnar batching & object storage 8:00 Jordan's Background (HFT trading firm, Google) 13:21 OSS vs proprietary infra 22:25 the future (or present) of engineering \w AI 29:03 the asymmetry of code review & testing (\w AI again) 33:49 jordan's StreamFusion project - accelerating Flink in Rust with DataFusion and Arrow 50:16 jordan's IceStream project - optimizing streaming upserts with a Paimon-backed index to async convert equality deletes 57:52 table formats (Apache Paimon, Iceberg & more) 1:11:35 object storage as the foundation of new infra
4
5
34
4,371
Use Grust, the universal fast graph API in Rust: pair.rs/grust Grust Book on First Pair Press: fp.ms/grust Foundation of the QueryGraph stack: QueryGraph.ai Supports 15 backends with GQL/Cypher, including the fastest and Rust graph databases such as @LakeSail — the AI native Rust Lakehouse @helixdb @SurrealDB @tursodatabase @evokoa_ai pgGraph Postgres 19 with graph extensions @falkordb @cocoindex_io for graph change streaming
Are we still talking loops or did we shift to graphs yet?
3
9
795
Everett Roeth retweeted
Spark accelerators like Photon speed up portions of execution, but are still tied to the JVM. Sail, written in Rust, removes the JVM entirely. The result is drastically better performance with no JVM tuning required. The same familiar Spark API on a Rust-native runtime. That’s Sail.
2
2
11
1,640