.On Mon, Oct 9, 2023 at 1:49 PM Reuven Lax <re...@google.com> wrote:
> Just FYI - the reason why names (including prefixes) in
> DynamicDestinations were parameterized via a lambda instead of just having
> the user add it via MapElements is performance. We discussed something
> along the lines of what you are suggesting (essentially having the user
> create a KV where the key contained the dynamic information). The problem
> was that often the size of the generated filepath was often much larger
> (sometimes by 2 OOM) than the information in the record, and there was a
> desire to avoid record blowup. e.g. the record might contain a single
> integer userid, and the filepath prefix would then be
> /long/path/to/output/users/<id>. This was especially bad in cases where the
> data had to be shuffled, and the existing dynamic destinations method
> allowed extracting the filepath only _after_ the shuffle.
>
That is a consideration I hadn't thought much of, thanks for bringing this
up.
> Now there may not be any good way to keep this benefit in a
> declarative approach such as YAML (or at least a good easy way - we could
> always allow the user to pass in a SQL expression to extract the filename
> from the record!), but we should keep in mind that this might mean that
> YAML-generated pipelines will be less efficient for certain use cases.
>
Yep, it's not as straightforward to do in a declarative way. I would like
to avoid mixing UDFs (with their associated languages and execution
environments) if possible. Though I'd like the performance of a
"straightforward" YAML pipeline to be that which one can get writing
straight-line Java (and possibly better, if we can leverage the structure
of schemas everywhere) this is not an absolute requirement for all
features.
I wonder if separating out a constant prefix vs. the dynamic stuff could be
sufficient to mitigate the blow-up of pre-computing this in most cases
(especially in the context of a larger pipeline). Alternatively, rather
than just a sharding pattern, one could have a full filepattern that
includes format parameters for dynamically computed bits as well as the
shard number, windowing info, etc. (There are pros and cons to this.)
> On Mon, Oct 9, 2023 at 12:37 PM Robert Bradshaw via dev <
> dev@beam.apache.org> wrote:
>
>> Currently the various file writing configurations take a single
>> parameter, path, which indicates where the (sharded) output should be
>> placed. In other words, one can write something like
>>
>> pipeline:
>> ...
>> sink:
>> type: WriteToParquet
>> config:
>> path: /beam/filesytem/dest
>>
>> and one gets files like "/beam/filesystem/dest-X-of-N"
>>
>> Of course, in practice file writing is often much more complicated than
>> this (especially when it comes to Streaming). For reference, I've included
>> links to our existing offerings in the various SDKs below. I'd like to
>> start a discussion about what else should go in the "config" parameter and
>> how it should be expressed in YAML.
>>
>> The primary concern is around naming. This can generally be split into
>> (1) the prefix, which must be provided by the users (2) the sharing
>> information, includes both shard counts (e.g. (the -X-of-N suffix) but also
>> windowing information (for streaming pipelines) which we may want to allow
>> the user to customize the formatting of, and (3) a suffix like .json or
>> .avro that is useful for both humans and tooling and can often be inferred
>> but should allow customization as well.
>>
>> An interesting case is that of dynamic destinations, where the prefix (or
>> other parameters) may themselves be functions of the records themselves. (I
>> am excluding the case where the format itself is variable--such cases are
>> probably better handled by explicitly partitioning the data and doing
>> multiple writes, as this introduces significant complexities and the set of
>> possible formats is generally finite and known ahead of time.) I propose
>> that we leverage the fact that we have structured data to be able to pull
>> out these dynamic parameters. For example, if we have an input data set
>> with a string column my_col we could allow something like
>>
>> config:
>> path: {dynamic: my_col}
>>
>> which would pull this information out at runtime. (With the MapToFields
>> transform, it is very easy to compute/append additional fields to existing
>> records.) Generally this field would then be stripped from the written
>> data, which would only see the subset of non-dynamically referenced columns
>> (though this could be configurable: we could add an attribute like
>> {dynamic: my_col, Keep: true} or require the set of columns to be actually
>> written (or elided) to be enumerated in the config or allow/require the
>> actual data to be written to be in a designated field of the "full" input
>> records as arranged by a preceding transform). It'd be great to get
>> input/impressions from a wide range of people here on what would be the
>> most natural. Often just writing out snippets of various alternatives can
>> be quite informative (though I'm avoiding putting them here for the moment
>> to avoid biasing ideas right off the bat).
>>
>> For streaming pipelines it is often essential to write data out in a
>> time-partitioned manner. The typical way to do this is to add the windowing
>> information into the shard specification itself, and a (set of) file(s) is
>> written on each window closing. Beam YAML already supports any transform
>> being given a "windowing" configuration which will cause a WindowInto
>> transform to be applied to its input(s) before application which can sit
>> naturally on a sink. We may want to consider if non-windowed writes make
>> sense as well (though how this interacts with the watermark and underlying
>> implementations are a large open question, so this is a larger change that
>> might make sense to defer).
>>
>> Note that I am explicitly excluding "coders" here. All data in YAML
>> should be schema'd, and writers should know how to write this structured
>> data. We may want to allow a "schema" field to allow a user to specify the
>> desired schema in a manner compatible with the sink format itself (e.g.
>> avro, json, whatever) that could be used both for validation and possibly
>> resolving ambiguities (e.g. if the sink has an enum format that is not
>> expressed in the schema of the input PCollection).
>>
>> Some other configuration options are that some formats (especially
>> text-based ones) allow for specification of an external compression type
>> (which may be inferable from the suffix), whether to write a single shard
>> if the input collection is empty or no shards at all (an occasional user
>> request that's supported for some Beam sinks now), whether to allow fixed
>> sharing (generally discouraged, as it disables things like automatic
>> shading based on input size, let alone dynamic work rebalancing, though
>> sometimes this is useful if the input is known to be small and a single
>> output is desired regardless of the restriction in parallelism), or other
>> sharding parameters (e.g. limiting the number of total elements or
>> (approximately) total number of bytes per output shard). Some of these
>> options may not be available/implemented for all formats--consideration
>> should be given as to how to handle this inconsistency (runtime errors for
>> unsupported combinations or simply not allowing them on any until all are
>> supported).
>>
>> A final consideration: we do not anticipate exposing the full complexity
>> of Beam in the YAML offering. For advanced users using a "real" SDK will
>> often be preferable, and we intend to provide a migration path from YAML to
>> a language of your choice (codegen) as a migration path. So we should
>> balance simplicity with completeness and utility here.
>>
>> Sure, we could just pick something, but given that the main point of YAML
>> is not capability, but expressibility and ease-of-use, I think it's worth
>> trying to get the expression of these concepts right. I'm sure many of you
>> have written a pipeline to files at some point in time; I'd welcome any
>> thoughts anyone has on the matter.
>>
>> - Robert
>>
>>
>> P.S. A related consideration: how should we consider the plain Read
>> (where that file pattern is given at pipeline construction) from the
>> ReadAll variants? Should they be separate transforms, or should we instead
>> allow the same named transform (e.g. ReadFromParquet) support both modes,
>> depending on whether an input PCollection or explicit file path is given
>> (the two being mutually exclusive, with exactly one required, and good
>> error messaging of course)?
>>
>>
>> Java:
>> https://beam.apache.org/releases/javadoc/current/org/apache/beam/sdk/io/TextIO.Write.html
>> Python:
>> https://beam.apache.org/releases/pydoc/current/apache_beam.io.textio.html#apache_beam.io.textio.WriteToText
>> Go:
>> https://pkg.go.dev/github.com/apache/beam/sdks/go/pkg/beam/io/textio#Write
>> Typescript:
>> https://beam.apache.org/releases/typedoc/current/functions/io_textio.writeToText.html
>> Scio:
>> https://spotify.github.io/scio/api/com/spotify/scio/io/TextIO$$WriteParam.html
>>
>>
