This is a really exciting development, thank you for putting together this proposal!
It looks like this thread and the linked GitHub issue has lots of input from folks who work with Arrow at a low level and have better familiarity with the Arrow specifications than I do, so I'll refrain from commenting on the technicalities of the proposal. I would, however, like to share my perspective as an application developer that heavily uses Arrow at higher levels for composing data systems. My main concern with the direction of this proposal is that it seems too narrowly focused on what the integration with DuckDB will look like (how the statistics can be fed into DuckDB). In many applications, executing the query is often the "last mile", and it's important to consider where the statistics will actually come from. To start, data might be sourced in various manners: - Arrow IPC files may be mapped from shared memory - Arrow IPC streams may be received via some RPC framework (à la Flight) - The Arrow libraries may be used to read from file formats like Parquet or CSV - ADBC drivers may be used to read from databases Note that in at least the first two cases, the system _executing the query_ will not be able to provide statistics simply because it is not actually the data producer. As an example, if Process A writes an Arrow IPC file to shared memory, and Process B wants to run a query on it -- how is Process B supposed to get the statistics for query planning? There are a few approaches that I anticipate application developers might consider: 1. Design an out-of-band mechanism for Process B to fetch statistics from Process A. 2. Design an encoding that is a superset of Arrow IPC and includes statistics information, allowing statistics to be communicated in-band. 3. Use custom schema metadata to communicate statistics in-band. Options 1 and 2 require considerably more effort than Option 3. Also, Option 3 feels somewhat natural because it makes sense for the statistics to come with the data (similar to how statistics are embedded in Parquet files). In some sense, the statistics actually *are* a property of the stream. In systems that I work on, we already use schema metadata to communicate information that is unrelated to the structure of the data. From my reading of the documentation [1], this sounds like a reasonable (and perhaps intended?) use of metadata, and nowhere is it mentioned that metadata must be used to determine schema equivalence. Unless there are other ways of producing stream-level application metadata outside of the schema/field metadata, the lack of purity was not a concern for me to begin with. I would appreciate an approach that communicates statistics via schema metadata, or at least in some in-band fashion that is consistent across the IPC and C data specifications. This would make it much easier to uniformly and transparently plumb statistics through applications, regardless of where they source Arrow data from. As developers are likely to create bespoke conventions for this anyways, it seems reasonable to standardize it as canonical metadata. I say this all as a happy user of DuckDB's Arrow scan functionality that is excited to see better query optimization capabilities. It's just that, in its current form, the changes in this proposal are not something I could foreseeably integrate with. Best, Shoumyo [1]: https://arrow.apache.org/docs/format/Columnar.html#custom-application-metadata From: dev@arrow.apache.org At: 05/23/24 10:10:51 UTC-4:00To: dev@arrow.apache.org Subject: Re: [DISCUSS] Statistics through the C data interface I want to +1 on what Dewey is saying here and some comments. Sutou Kouhei wrote: > ADBC may be a bit larger to use only for transmitting statistics. ADBC has statistics related APIs but it has more other APIs. It's impossible to keep the responsibility of communication protocols cleanly separated, but IMO, we should strive to keep the C Data Interface more of a Transport Protocol than an Application Protocol. Statistics are application dependent and can complicate the implementation of importers/exporters which would hinder the adoption of the C Data Interface. Statistics also bring in security concerns that are application-specific. e.g. can an algorithm trust min/max stats and risk producing incorrect results if the statistics are incorrect? A question that can't really be answered at the C Data Interface level. The need for more sophisticated statistics only grows with time, so there is no such thing as a "simple statistics schema". Protocols that produce/consume statistics might want to use the C Data Interface as a primitive for passing Arrow arrays of statistics. ADBC might be too big of a leap in complexity now, but "we just need C Data Interface + statistics" is unlikely to remain true for very long as projects grow in complexity. -- Felipe On Thu, May 23, 2024 at 9:57 AM Dewey Dunnington <de...@voltrondata.com.invalid> wrote: > > Thank you for the background! I understand that these statistics are > important for query planning; however, I am not sure that I follow why > we are constrained to the ArrowSchema to represent them. The examples > given seem to going through Python...would it be easier to request > statistics at a higher level of abstraction? There would already need > to be a separate mechanism to request an ArrowArrayStream with > statistics (unless the PyCapsule `requested_schema` argument would > suffice). > > > ADBC may be a bit larger to use only for transmitting > > statistics. ADBC has statistics related APIs but it has more > > other APIs. > > Some examples of producers given in the linked threads (Delta Lake, > Arrow Dataset) are well-suited to being wrapped by an ADBC driver. One > can implement an ADBC driver without defining all the methods (where > the producer could call AdbcConnectionGetStatistics(), although > AdbcStatementGetStatistics() might be more relevant here and doesn't > exist). One example listed (using an Arrow Table as a source) seems a > bit light to wrap in an ADBC driver; however, it would not take much > code to do so and the overhead of getting the reader via ADBC it is > something like 100 microseconds (tested via the ADBC R package's > "monkey driver" which wraps an existing stream as a statement). In any > case, the bulk of the code is building the statistics array. > > > How about the following schema for the > > statistics ArrowArray? It's based on ADBC. > > Whatever format for statistics is decided on, I imagine it should be > exactly the same as the ADBC standard? (Perhaps pushing changes > upstream if needed?). > > On Thu, May 23, 2024 at 3:21 AM Sutou Kouhei <k...@clear-code.com> wrote: > > > > Hi, > > > > > Why not simply pass the statistics ArrowArray separately in your > > > producer API of choice > > > > It seems that we should use the approach because all > > feedback said so. How about the following schema for the > > statistics ArrowArray? It's based on ADBC. > > > > | Field Name | Field Type | Comments | > > |--------------------------|-----------------------| -------- | > > | column_name | utf8 | (1) | > > | statistic_key | utf8 not null | (2) | > > | statistic_value | VALUE_SCHEMA not null | | > > | statistic_is_approximate | bool not null | (3) | > > > > 1. If null, then the statistic applies to the entire table. > > It's for "row_count". > > 2. We'll provide pre-defined keys such as "max", "min", > > "byte_width" and "distinct_count" but users can also use > > application specific keys. > > 3. If true, then the value is approximate or best-effort. > > > > VALUE_SCHEMA is a dense union with members: > > > > | Field Name | Field Type | > > |------------|------------| > > | int64 | int64 | > > | uint64 | uint64 | > > | float64 | float64 | > > | binary | binary | > > > > If a column is an int32 column, it uses int64 for > > "max"/"min". We don't provide all types here. Users should > > use a compatible type (int64 for a int32 column) instead. > > > > > > Thanks, > > -- > > kou > > > > In <a3ce5e96-176c-4226-9d74-6a458317a...@python.org> > > "Re: [DISCUSS] Statistics through the C data interface" on Wed, 22 May 2024 17:04:57 +0200, > > Antoine Pitrou <anto...@python.org> wrote: > > > > > > > > Hi Kou, > > > > > > I agree that Dewey that this is overstretching the capabilities of the > > > C Data Interface. In particular, stuffing a pointer as metadata value > > > and decreeing it immortal doesn't sound like a good design decision. > > > > > > Why not simply pass the statistics ArrowArray separately in your > > > producer API of choice (Dewey mentioned ADBC but it is of course just > > > a possible API among others)? > > > > > > Regards > > > > > > Antoine. > > > > > > > > > Le 22/05/2024 à 04:37, Sutou Kouhei a écrit : > > >> Hi, > > >> We're discussing how to provide statistics through the C > > >> data interface at: > > >> https://github.com/apache/arrow/issues/38837 > > >> If you're interested in this feature, could you share your > > >> comments? > > >> Motivation: > > >> We can interchange Apache Arrow data by the C data interface > > >> in the same process. For example, we can pass Apache Arrow > > >> data read by Apache Arrow C++ (provider) to DuckDB > > >> (consumer) through the C data interface. > > >> A provider may know Apache Arrow data statistics. For > > >> example, a provider can know statistics when it reads Apache > > >> Parquet data because Apache Parquet may provide statistics. > > >> But a consumer can't know statistics that are known by a > > >> producer. Because there isn't a standard way to provide > > >> statistics through the C data interface. If a consumer can > > >> know statistics, it can process Apache Arrow data faster > > >> based on statistics. > > >> Proposal: > > >> https://github.com/apache/arrow/issues/38837#issuecomment-2123728784 > > >> How about providing statistics as a metadata in ArrowSchema? > > >> We reserve "ARROW" namespace for internal Apache Arrow use: > > >> https://arrow.apache.org/docs/format/Columnar.html#custom-application-metadata > > >> > > >>> The ARROW pattern is a reserved namespace for internal > > >>> Arrow use in the custom_metadata fields. For example, > > >>> ARROW:extension:name. > > >> So we can use "ARROW:statistics" for the metadata key. > > >> We can represent statistics as a ArrowArray like ADBC does. > > >> Here is an example ArrowSchema that is for a record batch > > >> that has "int32 column1" and "string column2": > > >> ArrowSchema { > > >> .format = "+siu", > > >> .metadata = { > > >> "ARROW:statistics" => ArrowArray*, /* table-level statistics such as > > >> row count */ > > >> }, > > >> .children = { > > >> ArrowSchema { > > >> .name = "column1", > > >> .format = "i", > > >> .metadata = { > > >> "ARROW:statistics" => ArrowArray*, /* column-level statistics such as > > >> count distinct */ > > >> }, > > >> }, > > >> ArrowSchema { > > >> .name = "column2", > > >> .format = "u", > > >> .metadata = { > > >> "ARROW:statistics" => ArrowArray*, /* column-level statistics such as > > >> count distinct */ > > >> }, > > >> }, > > >> }, > > >> } > > >> The metadata value (ArrowArray* part) of '"ARROW:statistics" > > >> => ArrowArray*' is a base 10 string of the address of the > > >> ArrowArray. Because we can use only string for metadata > > >> value. You can't release the statistics ArrowArray*. (Its > > >> release is a no-op function.) It follows > > >> https://arrow.apache.org/docs/format/CDataInterface.html#member-allocation > > >> semantics. (The base ArrowSchema owns statistics > > >> ArrowArray*.) > > >> ArrowArray* for statistics use the following schema: > > >> | Field Name | Field Type | Comments | > > >> |----------------|----------------------------------| -------- | > > >> | key | string not null | (1) | > > >> | value | `VALUE_SCHEMA` not null | | > > >> | is_approximate | bool not null | (2) | > > >> 1. We'll provide pre-defined keys such as "max", "min", > > >> "byte_width" and "distinct_count" but users can also use > > >> application specific keys. > > >> 2. If true, then the value is approximate or best-effort. > > >> VALUE_SCHEMA is a dense union with members: > > >> | Field Name | Field Type | Comments | > > >> |------------|----------------------------------| -------- | > > >> | int64 | int64 | | > > >> | uint64 | uint64 | | > > >> | float64 | float64 | | > > >> | value | The same type of the ArrowSchema | (3) | > > >> | | that is belonged to. | | > > >> 3. If the ArrowSchema's type is string, this type is also string. > > >> TODO: Is "value" good name? If we refer it from the > > >> top-level statistics schema, we need to use > > >> "value.value". It's a bit strange... > > >> What do you think about this proposal? Could you share your > > >> comments? > > >> Thanks,
