Split up Arrow Crate

[tustvold]

TLDR rather than fighting entropy lets just brute-force compilation

Is your feature request related to a problem or challenge? Please describe what you are trying to do.

The arrow crate is getting rather large, and is starting to show up as a non-trivial bottleneck when compiling code, see #2170. There have been some efforts to reduce the amount of generated code, see #1858, but this is going to be a perpetual losing battle against new feature additions.

I think there are a couple of problems currently:

1. Limited build parallelism, especially if codegen-units is set low
2. Upstream crates have to "depend" on functionality they don't need, e.g. parquet depending on compute kernels
3. Minor changes force large amounts of recompilation, with incremental compilation only helping marginally
4. Codegen is rarely linear in complexity, consequently larger codegen units take longer than the same amount of code in smaller units

All these conspire to often result in an arrow shaped hole in compilation, where CPUs are left idle.

Some numbers from my local machine

• Release with default features: 232 seconds
• Release with default features without comparison kernels: 150 seconds
• Release with default features without compute kernels: 70 seconds
• Release without default features without compute kernels: 60 seconds

The vast majority of the time all bar a single core is idle.

Describe the solution you'd like

I would like to propose we split up the arrow crate, into a number of sub-crates that are then re-exported by the top-level arrow crate. Users can then choose to depend on the batteries included arrow crate, or more granular crates.

Initially I would propose the following split:

• arrow-csv: CSV reader support
• arrow-ipc: IPC support
• arrow-json: JSON support (related to Make JSON support Optional via Feature Flag #2300)
• arrow-compute: contents of compute module
• arrow-test: arrow test_utils (not published)
• arrow-core: everything else

There is definitely scope for splitting up the crates further after this, in particular the comparison kernels might be a good candidate to live on their own, but I think lets start small and go from there. I suspect there is a fair amount of disentangling that will be necessary to achieve this.

Describe alternatives you've considered

Feature flags are another way this can be handled, however, they have a couple of limitations:

• It is impractical to test the full combinatorial explosion of combinations, which allows for bugs to sneak through
• They are unified for a target which limits build parallelism, just because say DataFusion depends on arrow with CSV support, shouldn't force the parquet crate to wait for this to compile before it can start compiling
• Poor UX:
  • Discoverability is limited, it can be hard to determine what features gate what functionality
  • Hard to determine if the feature flag set is minimal, no equivalent of cargo-udeps
  • It can be a non-trivial detective exercise to determine why a given feature is being enabled
  • Necessitate counter-intuitive hacks to play nicely in multi-crate workspaces - see workspace hack

Additional context

@jimexist recently drove an initiative to do something similar to DataFusion which has worked very well - apache/datafusion#1750

FYI @alamb @jhorstmann @nevi-me

[alamb]

I like this idea, for what it is worth. 👍

[tustvold]

I've started work on this with #2693, I think the final split will likely end up being a different from what was initially proposed based on what components can easily be separated. I next plan to split out array data, followed by the arrays themselves. This should then allow splitting out some of the heavier kernels, e.g. sort, compare, cast, etc...

[alamb]

If the history with the datafusion split is any indication, this work is likely to end up generating lots of PRs

You can see how @jimexist broke that down and we tracked it in apache/datafusion#1750 -- perhaps something similar could be applied here.

I am happy to try and review these mechanical PRs more quickly so we can get the project done more quickly

[andygrove]

How feasible would it be to move the type definitions (such as DataType, Field, and Schema) into an arrow-types crate? Maybe this is a path to arrow2 being able to use a common type system.

[tustvold]

I don't see an obvious reason why that would not be possible, I'm not sure how generally useful the types will be without the array definitions though...

[andygrove]

The datafusion-sql crate only uses arrow::datatypes for example, so just depending on arrow-types there would presumably help with compilation times.

[andygrove]

@jorgecarleitao Would arrow2 and its ecosystem benefit from having an arrow-types crate as discussed here?

[jorgecarleitao]

Hey, Thanks for the ping!

I think it would not benefit arrow2 directly right now as it has different declarations for Field (e.g. we do not have dict_id on it). Arrow2 also has an extension (DataType::Extension).

With that said, imo it is still a good design - there are systems that only require DataType, Field, Schema, and functionality to read them from a file. One example is a data catalog based on arrow logical types.

I think that datafusion's logical plans could also only depend on types, but I could be wrong (it depends on how List scalars are represented there?).

[alamb]

I think that datafusion's logical plans could also only depend on types, but I could be wrong (it depends on how List scalars are represented there?).

List scalars are represented as Vec<Box<Field> so I agree it may be the case the logical plans could depend only on type

https://github.com/apache/arrow-datafusion/blob/d16457a0ba129b077935078e5cf89d028f598e0b/datafusion/common/src/scalar.rs#L81

[tustvold]

I've created #2711 which splits out the schema definitions into a crate called arrow-schema. I thought this was more clearly the logical types than something called arrow-types. PTAL 😄

[maxburke]

As a downstream user of Arrow, one of the things we find is that we need to fork Arrow-ecosystem crates to quickly integrate patches for missing features or bugs and I think one thing that I'm dreading is having to do that with an exploded Arrow crate, having to fork half a dozen Arrow packages, Parquet, all of the Datafusion crates, ..., seems like it'll be a royal pain.

[tustvold]

Hi @maxburke, the intention is to follow the work that was already performed for DataFusion, and would not involve splitting the repository. So I think you shouldn't have to maintain any more or less forks following this? The [patch.crates-io] directive would need to be for every crate though

[alamb]

I wonder is it done yet 🙏

[tustvold]

arrow-row and arrow-arith then yes, will likely do tomorrow

[alamb]