spark-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From James Baker <j.ba...@outlook.com>
Subject Re: [VOTE] [SPIP] SPARK-15689: Data Source API V2
Date Wed, 30 Aug 2017 21:57:11 GMT
I guess I was more suggesting that by coding up the powerful mode as the API, it becomes easy
for someone to layer an easy mode beneath it to enable simpler datasources to be integrated
(and that simple mode should be the out of scope thing).

Taking a small step back here, one of the places where I think I'm missing some context is
in understanding the target consumers of these interfaces. I've done some amount (though likely
not enough) of research about the places where people have had issues of API surface in the
past - the concrete tickets I've seen have been based on Cassandra integration where you want
to indicate clustering, and SAP HANA where they want to push down more complicated queries
through Spark. This proposal supports the former, but the amount of change required to support
clustering in the current API is not obviously high - whilst the current proposal for V2 seems
to make it very difficult to add support for pushing down plenty of aggregations in the future
(I've found the question of how to add GROUP BY to be pretty tricky to answer for the current
proposal).

Googling around for implementations of the current PrunedFilteredScan, I basically find a
lot of databases, which seems reasonable - SAP HANA, ElasticSearch, Solr, MongoDB, Apache
Phoenix, etc. I've talked to people who've used (some of) these connectors and the sticking
point has generally been that Spark needs to load a lot of data out in order to solve aggregations
that can be very efficiently pushed down into the datasources.

So, with this proposal it appears that we're optimising towards making it easy to write one-off
datasource integrations, with some amount of pluggability for people who want to do more complicated
things (the most interesting being bucketing integration). However, my guess is that this
isn't what the current major integrations suffer from; they suffer mostly from restrictions
in what they can push down (which broadly speaking are not going to go away).

So the place where I'm confused is that the current integrations can be made incrementally
better as a consequence of this, but the backing data systems have the features which enable
a step change which this API makes harder to achieve in the future. Who are the group of users
who benefit the most as a consequence of this change, like, who is the target consumer here?
My personal slant is that it's more important to improve support for other datastores than
it is to lower the barrier of entry - this is why I've been pushing here.

James

On Wed, 30 Aug 2017 at 09:37 Ryan Blue <rblue@netflix.com<mailto:rblue@netflix.com>>
wrote:

-1 (non-binding)

Sometimes it takes a VOTE thread to get people to actually read and comment, so thanks for
starting this one… but there’s still discussion happening on the prototype API, which
it hasn’t been updated. I’d like to see the proposal shaped by the ongoing discussion
so that we have a better, more concrete plan. I think that’s going to produces a better
SPIP.

The second reason for -1 is that I think the read- and write-side proposals should be separated.
The PR<https://github.com/cloud-fan/spark/pull/10> currently has “write path” listed
as a TODO item and most of the discussion I’ve seen is on the read side. I think it would
be better to separate the read and write APIs so we can focus on them individually.

An example of why we should focus on the write path separately is that the proposal says this:

Ideally partitioning/bucketing concept should not be exposed in the Data Source API V2, because
they are just techniques for data skipping and pre-partitioning. However, these 2 concepts
are already widely used in Spark, e.g. DataFrameWriter.partitionBy and DDL syntax like ADD
PARTITION. To be consistent, we need to add partitioning/bucketing to Data Source V2 . . .

Essentially, the some APIs mix DDL and DML operations. I’d like to consider ways to fix
that problem instead of carrying the problem forward to Data Source V2. We can solve this
by adding a high-level API for DDL and a better write/insert API that works well with it.
Clearly, that discussion is independent of the read path, which is why I think separating
the two proposals would be a win.

rb

​

On Wed, Aug 30, 2017 at 4:28 AM, Reynold Xin <rxin@databricks.com<mailto:rxin@databricks.com>>
wrote:
That might be good to do, but seems like orthogonal to this effort itself. It would be a completely
different interface.

On Wed, Aug 30, 2017 at 1:10 PM Wenchen Fan <cloud0fan@gmail.com<mailto:cloud0fan@gmail.com>>
wrote:
OK I agree with it, how about we add a new interface to push down the query plan, based on
the current framework? We can mark the query-plan-push-down interface as unstable, to save
the effort of designing a stable representation of query plan and maintaining forward compatibility.

On Wed, Aug 30, 2017 at 10:53 AM, James Baker <j.baker@outlook.com<mailto:j.baker@outlook.com>>
wrote:
I'll just focus on the one-by-one thing for now - it's the thing that blocks me the most.

I think the place where we're most confused here is on the cost of determining whether I can
push down a filter. For me, in order to work out whether I can push down a filter or satisfy
a sort, I might have to read plenty of data. That said, it's worth me doing this because I
can use this information to avoid reading >>that much data.

If you give me all the orderings, I will have to read that data many times (we stream it to
avoid keeping it in memory).

There's also a thing where our typical use cases have many filters (20+ is common). So, it's
likely not going to work to pass us all the combinations. That said, if I can tell you a cost,
I know what optimal looks like, why can't I just pick that myself?

The current design is friendly to simple datasources, but does not have the potential to support
this.

So the main problem we have with datasources v1 is that it's essentially impossible to leverage
a bunch of Spark features - I don't get to use bucketing or row batches or all the nice things
that I really want to use to get decent performance. Provided I can leverage these in a moderately
supported way which won't break in any given commit, I'll be pretty happy with anything that
lets me opt out of the restrictions.

My suggestion here is that if you make a mode which works well for complicated use cases,
you end up being able to write simple mode in terms of it very easily. So we could actually
provide two APIs, one that lets people who have more interesting datasources leverage the
cool Spark features, and one that lets people who just want to implement basic features do
that - I'd try to include some kind of layering here. I could probably sketch out something
here if that'd be useful?

James

On Tue, 29 Aug 2017 at 18:59 Wenchen Fan <cloud0fan@gmail.com<mailto:cloud0fan@gmail.com>>
wrote:
Hi James,

Thanks for your feedback! I think your concerns are all valid, but we need to make a tradeoff
here.

> Explicitly here, what I'm looking for is a convenient mechanism to accept a fully specified
set of arguments

The problem with this approach is: 1) if we wanna add more arguments in the future, it's really
hard to do without changing the existing interface. 2) if a user wants to implement a very
simple data source, he has to look at all the arguments and understand them, which may be
a burden for him.
I don't have a solution to solve these 2 problems, comments are welcome.


> There are loads of cases like this - you can imagine someone being able to push down
a sort before a filter is applied, but not afterwards. However, maybe the filter is so selective
that it's better to push down the filter and not handle the sort. I don't get to make this
decision, Spark does (but doesn't have good enough information to do it properly, whilst I
do). I want to be able to choose the parts I push down given knowledge of my datasource -
as defined the APIs don't let me do that, they're strictly more restrictive than the V1 APIs
in this way.

This is true, the current framework applies push downs one by one, incrementally. If a data
source wanna go back to accept a sort push down after it accepts a filter push down, it's
impossible with the current data source V2.
Fortunately, we have a solution for this problem. At Spark side, actually we do have a fully
specified set of arguments waiting to be pushed down, but Spark doesn't know which is the
best order to push them into data source. Spark can try every combination and ask the data
source to report a cost, then Spark can pick the best combination with the lowest cost. This
can also be implemented as a cost report interface, so that advanced data source can implement
it for optimal performance, and simple data source doesn't need to care about it and keep
simple.


The current design is very friendly to simple data source, and has the potential to support
complex data source, I prefer the current design over the plan push down one. What do you
think?


On Wed, Aug 30, 2017 at 5:53 AM, James Baker <j.baker@outlook.com<mailto:j.baker@outlook.com>>
wrote:
Yeah, for sure.

With the stable representation - agree that in the general case this is pretty intractable,
it restricts the modifications that you can do in the future too much. That said, it shouldn't
be as hard if you restrict yourself to the parts of the plan which are supported by the datasources
V2 API (which after all, need to be translateable properly into the future to support the
mixins proposed). This should have a pretty small scope in comparison. As long as the user
can bail out of nodes they don't understand, they should be ok, right?

That said, what would also be fine for us is a place to plug into an unstable query plan.

Explicitly here, what I'm looking for is a convenient mechanism to accept a fully specified
set of arguments (of which I can choose to ignore some), and return the information as to
which of them I'm ignoring. Taking a query plan of sorts is a way of doing this which IMO
is intuitive to the user. It also provides a convenient location to plug in things like stats.
Not at all married to the idea of using a query plan here; it just seemed convenient.

Regarding the users who just want to be able to pump data into Spark, my understanding is
that replacing isolated nodes in a query plan is easy. That said, our goal here is to be able
to push down as much as possible into the underlying datastore.

To your second question:

The issue is that if you build up pushdowns incrementally and not all at once, you end up
having to reject pushdowns and filters that you actually can do, which unnecessarily increases
overheads.

For example, the dataset

a b c
1 2 3
1 3 3
1 3 4
2 1 1
2 0 1

can efficiently push down sort(b, c) if I have already applied the filter a = 1, but otherwise
will force a sort in Spark. On the PR I detail a case I see where I can push down two equality
filters iff I am given them at the same time, whilst not being able to one at a time.

There are loads of cases like this - you can imagine someone being able to push down a sort
before a filter is applied, but not afterwards. However, maybe the filter is so selective
that it's better to push down the filter and not handle the sort. I don't get to make this
decision, Spark does (but doesn't have good enough information to do it properly, whilst I
do). I want to be able to choose the parts I push down given knowledge of my datasource -
as defined the APIs don't let me do that, they're strictly more restrictive than the V1 APIs
in this way.

The pattern of not considering things that can be done in bulk bites us in other ways. The
retrieval methods end up being trickier to implement than is necessary because frequently
a single operation provides the result of many of the getters, but the state is mutable, so
you end up with odd caches.

For example, the work I need to do to answer unhandledFilters in V1 is roughly the same as
the work I need to do to buildScan, so I want to cache it. This means that I end up with code
that looks like:

public final class CachingFoo implements Foo {
    private final Foo delegate;

    private List<Filter> currentFilters = emptyList();
    private Supplier<Bar> barSupplier = newSupplier(currentFilters);

    public CachingFoo(Foo delegate) {
        this.delegate = delegate;
    }

    private Supplier<Bar> newSupplier(List<Filter> filters) {
        return Suppliers.memoize(() -> delegate.computeBar(filters));
    }

    @Override
    public Bar computeBar(List<Filter> filters) {
        if (!filters.equals(currentFilters)) {
            currentFilters = filters;
            barSupplier = newSupplier(filters);
        }

        return barSupplier.get();
    }
}

which caches the result required in unhandledFilters on the expectation that Spark will call
buildScan afterwards and get to use the result..

This kind of cache becomes more prominent, but harder to deal with in the new APIs. As one
example here, the state I will need in order to compute accurate column stats internally will
likely be a subset of the work required in order to get the read tasks, tell you if I can
handle filters, etc, so I'll want to cache them for reuse. However, the cached information
needs to be appropriately invalidated when I add a new filter or sort order or limit, and
this makes implementing the APIs harder and more error-prone.

One thing that'd be great is a defined contract of the order in which Spark calls the methods
on your datasource (ideally this contract could be implied by the way the Java class structure
works, but otherwise I can just throw).

James

On Tue, 29 Aug 2017 at 02:56 Reynold Xin <rxin@databricks.com<mailto:rxin@databricks.com>>
wrote:
James,

Thanks for the comment. I think you just pointed out a trade-off between expressiveness and
API simplicity, compatibility and evolvability. For the max expressiveness, we'd want the
ability to expose full query plans, and let the data source decide which part of the query
plan can be pushed down.

The downside to that (full query plan push down) are:

1. It is extremely difficult to design a stable representation for logical / physical plan.
It is doable, but we'd be the first to do it. I'm not sure of any mainstream databases being
able to do that in the past. The design of that API itself, to make sure we have a good story
for backward and forward compatibility, would probably take months if not years. It might
still be good to do, or offer an experimental trait without compatibility guarantee that uses
the current Catalyst internal logical plan.

2. Most data source developers simply want a way to offer some data, without any pushdown.
Having to understand query plans is a burden rather than a gift.


Re: your point about the proposed v2 being worse than v1 for your use case.

Can you say more? You used the argument that in v2 there are more support for broader pushdown
and as a result it is harder to implement. That's how it is supposed to be. If a data source
simply implements one of the trait, it'd be logically identical to v1. I don't see why it
would be worse or better, other than v2 provides much stronger forward compatibility guarantees
than v1.


On Tue, Aug 29, 2017 at 4:54 AM, James Baker <j.baker@outlook.com<mailto:j.baker@outlook.com>>
wrote:
Copying from the code review comments I just submitted on the draft API (https://github.com/cloud-fan/spark/pull/10#pullrequestreview-59088745):

Context here is that I've spent some time implementing a Spark datasource and have had some
issues with the current API which are made worse in V2.

The general conclusion I’ve come to here is that this is very hard to actually implement
(in a similar but more aggressive way than DataSource V1, because of the extra methods and
dimensions we get in V2).

In DataSources V1 PrunedFilteredScan, the issue is that you are passed in the filters with
the buildScan method, and then passed in again with the unhandledFilters method.

However, the filters that you can’t handle might be data dependent, which the current API
does not handle well. Suppose I can handle filter A some of the time, and filter B some of
the time. If I’m passed in both, then either A and B are unhandled, or A, or B, or neither.
The work I have to do to work this out is essentially the same as I have to do while actually
generating my RDD (essentially I have to generate my partitions), so I end up doing some weird
caching work.

This V2 API proposal has the same issues, but perhaps moreso. In PrunedFilteredScan, there
is essentially one degree of freedom for pruning (filters), so you just have to implement
caching between unhandledFilters and buildScan. However, here we have many degrees of freedom;
sorts, individual filters, clustering, sampling, maybe aggregations eventually - and these
operations are not all commutative, and computing my support one-by-one can easily end up
being more expensive than computing all in one go.

For some trivial examples:

- After filtering, I might be sorted, whilst before filtering I might not be.

- Filtering with certain filters might affect my ability to push down others.

- Filtering with aggregations (as mooted) might not be possible to push down.

And with the API as currently mooted, I need to be able to go back and change my results because
they might change later.

Really what would be good here is to pass all of the filters and sorts etc all at once, and
then I return the parts I can’t handle.

I’d prefer in general that this be implemented by passing some kind of query plan to the
datasource which enables this kind of replacement. Explicitly don’t want to give the whole
query plan - that sounds painful - would prefer we push down only the parts of the query plan
we deem to be stable. With the mix-in approach, I don’t think we can guarantee the properties
we want without a two-phase thing - I’d really love to be able to just define a straightforward
union type which is our supported pushdown stuff, and then the user can transform and return
it.

I think this ends up being a more elegant API for consumers, and also far more intuitive.

James

On Mon, 28 Aug 2017 at 18:00 蒋星博 <jiangxb1987@gmail.com<mailto:jiangxb1987@gmail.com>>
wrote:
+1 (Non-binding)

Xiao Li <gatorsmile@gmail.com<mailto:gatorsmile@gmail.com>>于2017年8月28日
周一下午5:38写道:
+1

2017-08-28 12:45 GMT-07:00 Cody Koeninger <cody@koeninger.org<mailto:cody@koeninger.org>>:
Just wanted to point out that because the jira isn't labeled SPIP, it
won't have shown up linked from

http://spark.apache.org/improvement-proposals.html

On Mon, Aug 28, 2017 at 2:20 PM, Wenchen Fan <cloud0fan@gmail.com<mailto:cloud0fan@gmail.com>>
wrote:
> Hi all,
>
> It has been almost 2 weeks since I proposed the data source V2 for
> discussion, and we already got some feedbacks on the JIRA ticket and the
> prototype PR, so I'd like to call for a vote.
>
> The full document of the Data Source API V2 is:
> https://docs.google.com/document/d/1n_vUVbF4KD3gxTmkNEon5qdQ-Z8qU5Frf6WMQZ6jJVM/edit
>
> Note that, this vote should focus on high-level design/framework, not
> specified APIs, as we can always change/improve specified APIs during
> development.
>
> The vote will be up for the next 72 hours. Please reply with your vote:
>
> +1: Yeah, let's go forward and implement the SPIP.
> +0: Don't really care.
> -1: I don't think this is a good idea because of the following technical
> reasons.
>
> Thanks!

---------------------------------------------------------------------
To unsubscribe e-mail: dev-unsubscribe@spark.apache.org<mailto:dev-unsubscribe@spark.apache.org>








--
Ryan Blue
Software Engineer
Netflix
Mime
View raw message