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

Re: [Discuss] Proposing FLIP-25 - Support User State TTL Natively in Flink


thanks for your reply Fabian, about the overhead of storing the key bytes twice, I think maybe the situation is even a bit worse, in general, it means that the total amount of data to be stored has doubled(for each key, we need to store two records, one for timer, one for state). This maybe a bit uncomfortable when the state backend is based on RocksDB, because the timers are living together with the other states in the same RocksDB instance, which means that with using TTL, the amount of the records in RocksDB has to be doubled, I'm afraid this may hurt its performance.

Concerning the approach to add a timestamp to each value, TBH, I didn't have a deep thought on it yet and also not sure about it...In general, it can be described as follows:

- We attach a TS for every state record.
- When getting the record, we check the TS to see if its outdated.
- For the records that we will never touch again, we use the compaction to remove them. maybe one day one compaction is enough.

Best, Sihua

On 05/16/2018 16:38,Fabian Hueske<fhueske@xxxxxxxxx> wrote:

Yes. IMO it makes sense to put the logic into the abstract base classes to share the implementation across different state backends and state primitives.

The overhead of storing the key twice is a valid concern, but I'm not sure about the approach to add a timestamp to each value.
How would we discard state then? Iterating always over all (or a range of) keys to check if their state should be expired?
That would only work efficiently if we relax the clean-up logic which could be a valid design decision.

Best, Fabian

2018-05-14 9:33 GMT+02:00 sihua zhou <summerleafs@xxxxxxx>:

Hi Fabian,
thanks you very much for the reply, just a alternative. Can we implement the TTL logical in `AbstractStateBackend` and `AbstractState`? A simplest way is to append the `ts` to the state's value? and we use the backend's `current time`(its also can be event time and process time) to judge whether the data is outdated? The pros is that:
- state is puly backed by state backend.
- for each key-value, we only need to store the one copy of key? (if we implement TTL base on timer, we need to store two copys of key, one for the timer and one for the keyed state)

What do you think?


On 05/14/2018 15:20,Fabian Hueske<fhueske@xxxxxxxxx> wrote:
Hi Sihua,

I think it makes sense to couple state TTL to the timer service. We'll need some kind of timers to expire state, so I think we should reuse components that we have instead of implementing another timer service.

Moreover, using the same timer service and using the public state APIs helps to have a consistent TTL behavior across different state backend.

Best, Fabian

2018-05-14 8:51 GMT+02:00 sihua zhou <summerleafs@xxxxxxx>:

Hi Bowen,
thanks for your doc! I left some comments on the doc, the main concerning is that it makes me feel like a coupling that the TTL need to depend on `timer`. Because I think the TTL is a property of the state, so it should be backed by the state backend. If we implement the TTL base on the timer, than it looks like a coupling... it makes me feel that the backend for state becomes `state backend` + `timer`. And in fact, IMO, even the `timer` should depend on `state backend` in theroy, it's a type of HeapState that scoped to the `key group`(not scoped to per key like the current keyed state).

And I found the doc is for exact TTL, I wonder if we can support a relax TTL that could provides a better performance. Because to me, the reason that I need TTL is just to prevent the state size growing infinitly(I believe I'm not the only one like this), so a relax version is enough, if there is a relax TTL which have a better performance, I would prefer that.

What do you think?


On 05/14/2018 14:31,Bowen Li<bowenli86@xxxxxxxxx> wrote:
Thank you, Fabian! I've created the FLIP-25 page

To continue our discussion points:
1. I see what you mean now. I totally agree. Since we don't completely know
it now, shall we experiment or prototype a little bit before deciding this?
2. -
3. Adding tags to timers is an option.

Another option I came up with recently, is like this: let *InternalTimerService
*maintains user timers and TTL timers separately. Implementation classes of
InternalTimerService should add two new collections of timers,  e.g.
and *Ttl*EventTimeTimersQueue for HeapInternalTimerService. Upon
InternalTimerService#onProcessingTime() and advanceWatermark(), they will
first iterate through ProcessingTimeTimers and EventTimeTimers (user
timers) and then through *Ttl*ProcessingTimeTimers and *Ttl*EventTimeTimers

(Ttl timers).

We'll also add the following new internal APIs to register Ttl timers:

public void registerTtlProcessingTimeTimer(N namespace, long time);

public void registerTtlEventTimeTimer(N namespace, long time);

The biggest advantage, compared to option 1, is that it doesn't impact
existing timer-related checkpoint/savepoint, restore and migrations.

What do you think?  And, any other Flink committers want to chime in for
ideas? I've also documented the above two discussion points to the FLIP


On Wed, May 2, 2018 at 5:36 AM, Fabian Hueske <fhueske@xxxxxxxxx> wrote:

Hi Bowen,

1. The motivation to keep the TTL logic outside of the state backend was
mainly to avoid state backend custom implementations. If we have a generic
approach that would work for all state backends, we could try to put the
logic into a base class like AbstractStateBackend. After all, state cleanup
is tightly related to the responsibilities of state backends.
2. -
3. You're right. We should first call the user code before cleaning up.
The main problem that I see right now is that we have to distinguish
between user and TTL timers. AFAIK, the timer service does not support
timer tags (or another method) to distinguish timers.

I've given you the permissions to create and edit wiki pages.

Best, Fabian

2018-04-30 7:47 GMT+02:00 Bowen Li <bowenli86@xxxxxxxxx>:

Thanks Fabian! Here're my comments inline, and let me know your thoughts.

1. Where should the TTL code reside? In the state backend or in the

I believe TTL code should not reside in state backend, because a critical
design is that TTL is independent of and transparent to state backends.

According to my current knowledge, I think it probably should live with
operators in flink-streaming-java.

2. How to get notified about state accesses? I guess this depends on 1.

You previously suggested using callbacks. I believe that's the right way
to do decoupling.

3. How to avoid conflicts of TTL timers and user timers?

User timers might always be invoked first? This is not urgent, shall we
bake it for more time and discuss it along the way?

Besides, I don't have access to create a FLIP page under
k+Improvement+Proposals. Can you grant me the proper access?



On Tue, Apr 24, 2018 at 2:40 AM, Fabian Hueske <fhueske@xxxxxxxxx> wrote:

Hi Bowen,

Thanks for updating the proposal. This looks pretty good (as I said
There are a few areas, that are not yet fully fleshed out:

1. Where should the TTL code reside? In the state backend or in the
2. How to get notified about state accesses? I guess this depends on 1.
3. How to avoid conflicts of TTL timers and user timers?

@Stefan (in CC) might have some ideas on these issues as well.

Cheers, Fabian

2018-04-22 21:14 GMT+02:00 Bowen <bowenli86@xxxxxxxxx>:

Hello community,

We've come up with a completely new design for Flink state TTL, documented


and have run it by a few Flink PMC/committers.

What do you think? We'd love to hear feedbacks from you


On Wed, Feb 7, 2018 at 12:53 PM, Fabian Hueske <fhueske@xxxxxxxxx>

Hi Bowen,

Thanks for the proposal! I think state TTL would be a great feature!
Actually, we have implemented this for SQL / Table API [1].
I've added a couple of comments to the design doc.

In principle, I'm not sure if this functionality should be added to the
state backends.
We could also use the existing timer service which would have a few
benefits (see my comments in the docs).

Best, Fabian


2018-02-06 8:26 GMT+01:00 Bowen Li <bowenli86@xxxxxxxxx>:

Hi guys,

I want to propose a new FLIP -- FLIP-25 - Support User State TTL
in Flink. This has been one of most handy and most frequently asked
features in Flink community. The jira ticket is FLINK-3089

I've written a rough design
and developed prototypes for both heap and rocksdb state backends.

My question is: shall we create a FLIP page for this? Can I be
granted the
privileges of creating pages in