[Feature Request] Liquid Clustering

[tdas]

Overview

We propose to introduce Liquid Clustering, a new effort to revamp how clustering works in Delta, which addresses the shortcomings of Hive-style partitioning and current ZORDER clustering.

Motivation

Partitioning/clustering is a common technique to manage datasets to reduce unnecessary data processing. Hive-style partitioning and ZORDER (multi-dimensional) clustering are existing solutions in Delta, but both have limitations.

Hive-style partitioning

Hive-style partitioning clusters data such that every file contains exactly one distinct combination of partition column values. Although Hive-style partitioning works well if tuned correctly, there are limitations:

• Since partition values are physical boundaries for files, Hive-style partitioning by high cardinality columns will create many small files that cannot be combined, resulting in poor scan performance.
• In Spark/Delta, once a table is partitioned, its partition strategy cannot be changed, thus being unable to adapt to new use cases such as query pattern changes, etc.

ZORDER clustering

ZORDER is a muti-dimensional clustering technique used in Delta. The OPTIMIZE ZORDER BY command applies ZORDER clustering and improves the performance of queries that utilize ZORDER BY columns in their predicates. However, it has the following limitations:

• Any new data ingested after the OPTIMIZE ZORDER BY run is not automatically clustered, and the user needs to rerun the command to cluster the new data.
• OPTIMIZE ZORDER BY reclusters already well-clustered data, resulting in high write amplification.
• ZORDER BY columns are not persisted and the user is required to remember the previous ZORDER BY columns, often causing user errors.

Detailed Design

Please refer to the document here.

[akshaysinghas]

Hi guys, do we have any white paper or link that I can refer to understand MDC, or liquid clustering? I know requirement doc is on the way but any other resource that can throw more light on this feature?

[felipepessoto]

This feature in on Delta Lake 3.0.0 Preview release notes, is it expected to have an 3.0.0 RC2 with the feature?

Usually, the Delta final version is released few weeks after RC1. It seems this won't make to the 3.0.0 final?

[kangkaisen]

@tdas Hi，does Liquid Clustering use the DBSCAN algorithm? Thanks.

[nachiketrajput-upl]

@akshaysinghUdaan: Kindly use the link, https://www.databricks.com/blog/announcing-delta-lake-30-new-universal-format-and-liquid-clustering

[loquisgon]

This needs a LOT more detail: "Liquid clustering uses a continuous fractal space filling curve as a multi-dimensional clustering technique, which significantly improves data skipping"

[loquisgon]

"For multidimensional databases, Hilbert order has been proposed to be used instead of Z order because it has better locality-preserving behavior. For example, Hilbert curves have been used to compress and accelerate R-tree indexes[9] (see Hilbert R-tree). They have also been used to help compress data warehouses.[10][11]"

https://en.wikipedia.org/wiki/Hilbert_curve#:~:text=The%20Hilbert%20curve%20(also%20known,by%20Giuseppe%20Peano%20in%201890.

[imback82]

This needs a LOT more detail: "Liquid clustering uses a continuous fractal space filling curve as a multi-dimensional clustering technique, which significantly improves data skipping"

@loquisgon we are working on the design doc and we will have more details on clustering algorithm. Stay tuned!

[KamilKandzia]

@imback82 Any progress on the doc or implementing solution in delta project? Databricks 13.2 has already implemented the liquid clustering, but we still have lack of information about this solution

[dabao521]

@vkorukanti can you assign this issue to me?

[zedtang]

Hello everyone, we've added the design doc to this issue. Please feel free to review it. Thank you!

[felipepessoto]

@zedtang if you don't mind to share the doc using this option which allows to print and comment (if you select Commenter). Even if you leave Viewer permission only, at least that option allow us to print it.

https://support.google.com/drive/answer/2494822?hl=en&co=GENIE.Platform%3DDesktop#zippy=%2Cshare-a-file-publicly

Previous design docs were shared that way

[zedtang]

Hi @felipepessoto , I updated the link to give everyone commenter access!

[zmeir]

Hi, will it be possible to use liquid clustering on partitioned tables?

[ghost]

In the design doc I don't see any info if clustering can be applied to existing Delta tables. As we use liquid from databricks already we observed that either you have to apply liquid to an empty table or you have to use the like command which basically copies and creates a new Delta table

I guess it will be really frustrating if we can't apply liquid to existing Delta tables :)

Thanks already everyone

[auckenth]

According to https://delta.io/pdfs/DLTDG_ER3.pdf (page 131) k-d trees are used to implement liquid clustering instead of space filling curves. Are there any pros and cons for an implementation using k-d trees?

[imback82]

In the design doc I don't see any info if clustering can be applied to existing Delta tables. As we use liquid from databricks already we observed that either you have to apply liquid to an empty table or you have to use the like command which basically copies and creates a new Delta table

I guess it will be really frustrating if we can't apply liquid to existing Delta tables :)

Good point. I think we should be able to migrate existing tables in place by ALTER TABLE ... CLUSTER BY .... This should be straightforward for un-partitioned tables, but it will be tricky to implement it for partitioned tables.

[dennyglee]

According to https://delta.io/pdfs/DLTDG_ER3.pdf (page 131) k-d trees are used to implement liquid clustering instead of space filling curves. Are there any pros and cons for an implementation using k-d trees?

Oh sorry about that @auckenth - this is an error within the book. We will be updating the book chapter with the latest info per the design doc.

[ghost]

In the design doc I don't see any info if clustering can be applied to existing Delta tables. As we use liquid from databricks already we observed that either you have to apply liquid to an empty table or you have to use the like command which basically copies and creates a new Delta table
I guess it will be really frustrating if we can't apply liquid to existing Delta tables :)

Good point. I think we should be able to migrate existing tables in place by ALTER TABLE ... CLUSTER BY .... This should be straightforward for un-partitioned tables, but it will be tricky to implement it for partitioned tables.

@imback82 This would be great. I think considering non-partitioned tables which are optionally z-ordered is enough as a baseline. For partitioned tables you neverless accept that any changes might need an overwrite.

[acruise]

It'd be nice if someone could address comments and questions in the design doc. :)

[adriangb]

Sorry if this is not the right place to ask, but is there anyway with liquid cluster to enforce some level of clustering? In particular I want to cluster by something like project,day and implement retention such that I delete data older than X days which is different for each project. With hive style I can partition on those columns and delete entire files without rewrites.

[zedtang]

It'd be nice if someone could address comments and questions in the design doc. :)

Sorry for the delay, I replied in the design doc.

[sezruby]

@zedtang Are you planning to support "cluster on write" ?

[zedtang]

@zedtang Are you planning to support "cluster on write" ?

Hi @sezruby , May I ask what your use case is for this?

[sezruby]

@zedtang With the feature, we don't need additional OPTIMIZE run after ingesting large data.
We may integrate it with optimize write feature for non-partitioned data.

[zedtang]

It's probably tricky to implement clustering on write. (1) clustering works well only when a sufficient amount of data has accumulated and can be clustered together, so doing on every write (commonly small writes) may not produce well-clustered data at the end, and one has to run offline optimize jobs anyway (2) clustering on write will require some kind of sorting, which will come at a cost to write latency. so it's not an easy decision.

It would be very cool if someone proposes a design and shows that it will work well for some workloads without sacrificing write latency/throughput significantly