Allowing an annotation on DoFn's that produce deterministic output could be added in the future, but doesn't seem like a great option.
1. It is a correctness issue to assume a DoFn is deterministic and be wrong, so we would need to assume all transform outputs are non-deterministic unless annotated. Getting this correct is difficult (for example, GBK is surprisingly non-deterministic except in specific cases). 2. It is unlikely to be a major performance improvement, given that any non-deterministic transform prior to a sink (which are most likely to require deterministic input) will cause additional work to be needed. Based on this, it seems like the risk of allowing an annotation is high while the potential for performance improvements is low. The current proposal (not allowing an annotation) makes sense for now, until we can demonstrate that the impact on performance is high in cases that could be avoided with an annotation (in real-world use). -- Ben On Tue, Mar 21, 2017 at 2:05 PM vikas rk <vikky...@gmail.com> wrote: +1 for the general idea of runners handling it over hard-coded implementation strategy. For the Write transform I believe you are talking about ApplyShardingKey < https://github.com/apache/beam/blob/d66029cafde152c0a46ebd276ddfa4c3e7fd3433/sdks/java/core/src/main/java/org/apache/beam/sdk/io/Write.java#L304 > which introduces non deterministic behavior when retried? *Let a DoFn declare (mechanism not important right now) that it "requiresdeterministic input"* *Each runner will need a way to induce deterministic input - the obviouschoice being a materialization.* Does this mean that a runner will always materialize (or whatever the strategy is) an input PCollection to this DoFn even though the PCollection might have been produced by deterministic transforms? Would it make sense to also let DoFns declare if they produce non-deterministic output? -Vikas On 21 March 2017 at 13:52, Stephen Sisk <s...@google.com.invalid> wrote: > Hey Kenn- > > this seems important, but I don't have all the context on what the problem > is. > > Can you explain this sentence "Specifically, there is pseudorandom data > generated and once it has been observed and used to produce a side effect, > it cannot be regenerated without erroneous results." ? > > Where is the pseudorandom data coming from? Perhaps a concrete example > would help? > > S > > > On Tue, Mar 21, 2017 at 1:22 PM Kenneth Knowles <k...@google.com.invalid> > wrote: > > > Problem: > > > > I will drop all nuance and say that the `Write` transform as it exists in > > the SDK is incorrect until we add some specification and APIs. We can't > > keep shipping an SDK with an unsafe transform in it, and IMO this > certainly > > blocks a stable release. > > > > Specifically, there is pseudorandom data generated and once it has been > > observed and used to produce a side effect, it cannot be regenerated > > without erroneous results. > > > > This generalizes: For some side-effecting user-defined functions, it is > > vital that even across retries/replays they have a consistent view of the > > contents of their input PCollection, because their effect on the outside > > world cannot be retracted if/when they fail and are retried. Once the > > runner ensures a consistent view of the input, it is then their own > > responsibility to be idempotent. > > > > Ideally we should specify this requirement for the user-defined function > > without imposing any particular implementation strategy on Beam runners. > > > > Proposal: > > > > 1. Let a DoFn declare (mechanism not important right now) that it > "requires > > deterministic input". > > > > 2. Each runner will need a way to induce deterministic input - the > obvious > > choice being a materialization. > > > > I want to keep the discussion focused, so I'm leaving out any > possibilities > > of taking this further. > > > > Regarding performance: Today places that require this tend to be already > > paying the cost via GroupByKey / Reshuffle operations, since that was a > > simple way to induce determinism in batch Dataflow* (doesn't work for > most > > other runners nor for streaming Dataflow). This change will replace a > > hard-coded implementation strategy with a requirement that may be > fulfilled > > in the most efficient way available. > > > > Thoughts? > > > > Kenn (w/ lots of consult from colleagues, especially Ben) > > > > * There is some overlap with the reshuffle/redistribute discussion > because > > of this historical situation, but I would like to leave that broader > > discussion out of this correctness issue. > > >
