[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: SplittableDoFn


Can you give details on what the synchronization is per? Is it per key, or global to each worker?

On Fri, Sep 21, 2018 at 2:10 PM Lukasz Cwik <lcwik@xxxxxxxxxx> wrote:
As I was looking at the SplittableDoFn API while working towards making a proposal for how the backlog/splitting API could look, I found some sharp edges that could be improved.

I noticed that:
1) We require users to write thread safe code, this is something that we haven't asked of users when writing a DoFn.
2) We "internal" methods within the RestrictionTracker that are not meant to be used by the runner.

I can fix these issues by giving the user a forwarding restriction tracker[1] that provides an appropriate level of synchronization as needed and also provides the necessary observation hooks to see when a claim failed or succeeded.

This requires a change to our experimental API since we need to pass a RestrictionTracker to the @ProcessElement method instead of a sub-type of RestrictionTracker.
@ProcessElement
processElement(ProcessContext context, OffsetRangeTracker tracker) { ... }
becomes:
@ProcessElement
processElement(ProcessContext context, RestrictionTracker<OffsetRange, Long> tracker) { ... }

This provides an additional benefit that it prevents users from working around the RestrictionTracker APIs and potentially making underlying changes to the tracker outside of the tryClaim call.

Full implementation is available within this PR[2] and was wondering what people thought.



On Mon, Sep 17, 2018 at 12:45 PM Lukasz Cwik <lcwik@xxxxxxxxxx> wrote:
The changes to the API have not been proposed yet. So far it has all been about what is the representation and why.

For splitting, the current idea has been about using the backlog as a way of telling the SplittableDoFn where to split, so it would be in terms of whatever the SDK decided to report.
The runner always chooses a number for backlog that is relative to the SDKs reported backlog. It would be upto the SDK to round/clamp the number given by the Runner to represent something meaningful for itself.
For example if the backlog that the SDK was reporting was bytes remaining in a file such as 500, then the Runner could provide some value like 212.2 which the SDK would then round to 212.
If the backlog that the SDK was reporting 57 pubsub messages, then the Runner could provide a value like 300 which would mean to read 57 values and then another 243 as part of the current restriction.

I believe that BoundedSource/UnboundedSource will have wrappers added that provide a basic SplittableDoFn implementation so existing IOs should be migrated over without API changes.

On Mon, Sep 17, 2018 at 1:11 AM Ismaël Mejía <iemejia@xxxxxxxxx> wrote:
Thanks a lot Luke for bringing this back to the mailing list and Ryan for taking
the notes.

I would like to know if there was some discussion, or if you guys have given
some thought to the required changes in the SDK (API) part. What will be the
equivalent of `splitAtFraction` and what should IO authors do to support it..

On Sat, Sep 15, 2018 at 1:52 AM Lukasz Cwik <lcwik@xxxxxxxxxx> wrote:
>
> Thanks to everyone who joined and for the questions asked.
>
> Ryan graciously collected notes of the discussion: https://docs.google.com/document/d/1kjJLGIiNAGvDiUCMEtQbw8tyOXESvwGeGZLL-0M06fQ/edit?usp=sharing
>
> The summary was that bringing BoundedSource/UnboundedSource into using a unified backlog-reporting mechanism with optional other signals that Dataflow has found useful (such as is the remaining restriction splittable (yes, no, unknown)). Other runners can use or not. SDFs should report backlog and watermark as minimum bar. The backlog should use an arbitrary precision float such as Java BigDecimal to prevent issues where limited precision removes the ability to compute delta efficiently.
>
>
>
> On Wed, Sep 12, 2018 at 3:54 PM Lukasz Cwik <lcwik@xxxxxxxxxx> wrote:
>>
>> Here is the link to join the discussion: https://meet.google.com/idc-japs-hwf
>> Remember that it is this Friday Sept 14th from 11am-noon PST.
>>
>>
>>
>> On Mon, Sep 10, 2018 at 7:30 AM Maximilian Michels <mxm@xxxxxxxxxx> wrote:
>>>
>>> Thanks for moving forward with this, Lukasz!
>>>
>>> Unfortunately, can't make it on Friday but I'll sync with somebody on
>>> the call (e.g. Ryan) about your discussion.
>>>
>>> On 08.09.18 02:00, Lukasz Cwik wrote:
>>> > Thanks for everyone who wanted to fill out the doodle poll. The most
>>> > popular time was Friday Sept 14th from 11am-noon PST. I'll send out a
>>> > calendar invite and meeting link early next week.
>>> >
>>> > I have received a lot of feedback on the document and have addressed
>>> > some parts of it including:
>>> > * clarifying terminology
>>> > * processing skew due to some restrictions having their watermarks much
>>> > further behind then others affecting scheduling of bundles by runners
>>> > * external throttling & I/O wait overhead reporting to make sure we
>>> > don't overscale
>>> >
>>> > Areas that still need additional feedback and details are:
>>> > * reporting progress around the work that is done and is active
>>> > * more examples
>>> > * unbounded restrictions being caused by an unbounded number of splits
>>> > of existing unbounded restrictions (infinite work growth)
>>> > * whether we should be reporting this information at the PTransform
>>> > level or at the bundle level
>>> >
>>> >
>>> >
>>> > On Wed, Sep 5, 2018 at 1:53 PM Lukasz Cwik <lcwik@xxxxxxxxxx
>>> > <mailto:lcwik@xxxxxxxxxx>> wrote:
>>> >
>>> >     Thanks to all those who have provided interest in this topic by the
>>> >     questions they have asked on the doc already and for those
>>> >     interested in having this discussion. I have setup this doodle to
>>> >     allow people to provide their availability:
>>> >     https://doodle.com/poll/nrw7w84255xnfwqy
>>> >
>>> >     I'll send out the chosen time based upon peoples availability and a
>>> >     Hangout link by end of day Friday so please mark your availability
>>> >     using the link above.
>>> >
>>> >     The agenda of the meeting will be as follows:
>>> >     * Overview of the proposal
>>> >     * Enumerate and discuss/answer questions brought up in the meeting
>>> >
>>> >     Note that all questions and any discussions/answers provided will be
>>> >     added to the doc for those who are unable to attend.
>>> >
>>> >     On Fri, Aug 31, 2018 at 9:47 AM Jean-Baptiste Onofré
>>> >     <jb@xxxxxxxxxxxx <mailto:jb@xxxxxxxxxxxx>> wrote:
>>> >
>>> >         +1
>>> >
>>> >         Regards
>>> >         JB
>>> >         Le 31 août 2018, à 18:22, Lukasz Cwik <lcwik@xxxxxxxxxx
>>> >         <mailto:lcwik@xxxxxxxxxx>> a écrit:
>>> >
>>> >             That is possible, I'll take people's date/time suggestions
>>> >             and create a simple online poll with them.
>>> >
>>> >             On Fri, Aug 31, 2018 at 2:22 AM Robert Bradshaw
>>> >             <robertwb@xxxxxxxxxx <mailto:robertwb@xxxxxxxxxx>> wrote:
>>> >
>>> >                 Thanks for taking this up. I added some comments to the
>>> >                 doc. A European-friendly time for discussion would
>>> >                 be great.
>>> >
>>> >                 On Fri, Aug 31, 2018 at 3:14 AM Lukasz Cwik
>>> >                 <lcwik@xxxxxxxxxx <mailto:lcwik@xxxxxxxxxx>> wrote:
>>> >
>>> >                     I came up with a proposal[1] for a progress model
>>> >                     solely based off of the backlog and that splits
>>> >                     should be based upon the remaining backlog we want
>>> >                     the SDK to split at. I also give recommendations to
>>> >                     runner authors as to how an autoscaling system could
>>> >                     work based upon the measured backlog. A lot of
>>> >                     discussions around progress reporting and splitting
>>> >                     in the past has always been around finding an
>>> >                     optimal solution, after reading a lot of information
>>> >                     about work stealing, I don't believe there is a
>>> >                     general solution and it really is upto
>>> >                     SplittableDoFns to be well behaved. I did not do
>>> >                     much work in classifying what a well behaved
>>> >                     SplittableDoFn is though. Much of this work builds
>>> >                     off ideas that Eugene had documented in the past[2].
>>> >
>>> >                     I could use the communities wide knowledge of
>>> >                     different I/Os to see if computing the backlog is
>>> >                     practical in the way that I'm suggesting and to
>>> >                     gather people's feedback.
>>> >
>>> >                     If there is a lot of interest, I would like to hold
>>> >                     a community video conference between Sept 10th and
>>> >                     14th about this topic. Please reply with your
>>> >                     availability by Sept 6th if your interested.
>>> >
>>> >                     1: https://s.apache.org/beam-bundles-backlog-splitting
>>> >                     2: https://s.apache.org/beam-breaking-fusion
>>> >
>>> >                     On Mon, Aug 13, 2018 at 10:21 AM Jean-Baptiste
>>> >                     Onofré <jb@xxxxxxxxxxxx <mailto:jb@xxxxxxxxxxxx>> wrote:
>>> >
>>> >                         Awesome !
>>> >
>>> >                         Thanks Luke !
>>> >
>>> >                         I plan to work with you and others on this one.
>>> >
>>> >                         Regards
>>> >                         JB
>>> >                         Le 13 août 2018, à 19:14, Lukasz Cwik
>>> >                         <lcwik@xxxxxxxxxx <mailto:lcwik@xxxxxxxxxx>> a
>>> >                         écrit:
>>> >
>>> >                             I wanted to reach out that I will be
>>> >                             continuing from where Eugene left off with
>>> >                             SplittableDoFn. I know that many of you have
>>> >                             done a bunch of work with IOs and/or runner
>>> >                             integration for SplittableDoFn and would
>>> >                             appreciate your help in advancing this
>>> >                             awesome idea. If you have questions or
>>> >                             things you want to get reviewed related to
>>> >                             SplittableDoFn, feel free to send them my
>>> >                             way or include me on anything SplittableDoFn
>>> >                             related.
>>> >
>>> >                             I was part of several discussions with
>>> >                             Eugene and I think the biggest outstanding
>>> >                             design portion is to figure out how dynamic
>>> >                             work rebalancing would play out with the
>>> >                             portability APIs. This includes reporting of
>>> >                             progress from within a bundle. I know that
>>> >                             Eugene had shared some documents in this
>>> >                             regard but the position / split models
>>> >                             didn't work too cleanly in a unified sense
>>> >                             for bounded and unbounded SplittableDoFns.
>>> >                             It will likely take me awhile to gather my
>>> >                             thoughts but could use your expertise as to
>>> >                             how compatible these ideas are with respect
>>> >                             to to IOs and runners
>>> >                             Flink/Spark/Dataflow/Samza/Apex/... and
>>> >                             obviously help during implementation.
>>> >