Hi, I would like to propose a modification to this FLIP.
Based on the feedback that we were receiving after publishing this document and during Flink Forward, I was growing more and more anxious about one issue here: having to persist all buffered in-flight data at once. As the volume of this data might be large (GBs per TaskManager even with small clusters and relatively simple jobs), the time to persist all of this data at once might be quite substantial. To address this issue, I would like to propose that at first we implement a variant of unaligned checkpoints, just as written down in FLIP-76, but with continuous spilling - all data will be persisted/spilled continuously, all the time as they come - not at once when the checkpoint starts. Think about this proposal as incremental way of persisting the data. Pros of continuous spilling: + faster checkpointing, as there will be no need to store GBs of data, just flush/close. + more predictable behaviour. Instead of jerky/varying/spike IO/CPU loads, steady records throughput and spilling. Cons of continuous spilling: - need to persist all of the network traffic instead of persisting just the in-flight data Larger volume of persisted data doesn’t matter that much from the perspective of the throughput, as if you are unable to spill the data faster than to process them, unaligned checkpoints are worse option compared to the aligned checkpoints [1]. If checkpoints are frequent it also doesn’t matter [2]. The true downside is if checkpoints are infrequent and you have to for example pay $ for the extra storage or extra network traffic to the storage. On the other hand, continuous spilling (persistent communication channels?) might have an added benefit of enabling us localised failures - failure of one node will not necessarily bring down the whole cluster. As I mentioned, I’m proposing to just start with the continuous spilling. It might be more costly in some scenarios, but it will offer the most stable and predictable performance with the lowest checkpoint latency. It’s not perfect, it won’t solve all of the use cases, but frankly all of the other options have their own blind spots, and continuous spilling should at least fully solve relatively low throughput use cases. We can later build on top of that solution, expanding it with the following features: 1. Do not spill continuously if there is no backpressure. For example provide a timeout: start spilling pre-emptively/continuously if some buffer was not processed within X seconds. 2. Start spilling only once the checkpoint starts (this is the exact proposal from the current FLIP-76). 3. Initially we want to spill to a Flink’s FileSystem (for example S3), but in the future we are considering other options, for example Apache Bookeeper. What do you think? Piotrek [1] I’m assuming that the spilling throughput per node can go up to ~30MB/s. If your Flink's job data processing rate is 100MB/s, spilling in-flight data will take 3.3 times longer than waiting for the alignment. On the other hand if data processing rate is 10MB/s, overhead of continuous spilling is relatively low. [2] With checkpoints every one minute, with data processing throughput 30MB/s per node, we would have to persist 1.8GB of data per node between the checkpoints, which is similar order of magnitude as buffered in-flight data under the back-pressure. With higher throughput, unaligned checkpoints are not helping ([1]). With lower throughput, both the original proposal and continuous spilling would have to effectively persist all of the data anyway. > On 10 Oct 2019, at 19:51, Yun Tang <myas...@live.com> wrote: > > Hi Arvid > > +1 for this future which has been hoped for a long time. End-to-end exactly > once job could benefit from quicker checkpoint completion. > > > Best > Yun Tang > ________________________________ > From: Yun Gao <yungao...@aliyun.com.INVALID> > Sent: Thursday, October 10, 2019 18:39 > To: dev <dev@flink.apache.org> > Subject: Re: [DISCUSS] FLIP-76: Unaligned checkpoints > > Hi Arvid, > > Very thanks for bring up the discussion! From our side unable to > finish the checkpoint is commonly met for online jobs, therefore +1 from my > side to implement this. > A tiny issue of the FLIP is that the Discussion Thread URL attached > seems to be not right. > > > Best, > Yun > > > ------------------------------------------------------------------ > From:Arvid Heise <ar...@ververica.com> > Send Time:2019 Sep. 30 (Mon.) 20:31 > To:dev <dev@flink.apache.org> > Subject:[DISCUSS] FLIP-76: Unaligned checkpoints > > Hi Devs, > > I would like to start the formal discussion about FLIP-76 [1], which > improves the checkpoint latency in systems under backpressure, where a > checkpoint can take hours to complete in the worst case. I recommend the > thread "checkpointing under backpressure" [2] to get a good idea why users > are not satisfied with the current behavior. The key points: > > - Since the checkpoint barrier flows much slower through the > back-pressured channels, the other channels and their upstream operators > are effectively blocked during checkpointing. > - The checkpoint barrier takes a long time to reach the sinks causing > long checkpointing times. A longer checkpointing time in turn means that > the checkpoint will be fairly outdated once done. Since a heavily utilized > pipeline is inherently more fragile, we may run into a vicious cycle of > late checkpoints, crash, recovery to a rather outdated checkpoint, more > back pressure, and even later checkpoints, which would result in little to > no progress in the application. > > The FLIP proposes "unaligned checkpoints" which improves the current state, > such that > > - Upstream processes can continue to produce data, even if some operator > still waits on a checkpoint barrier on a specific input channel. > - Checkpointing times are heavily reduced across the execution graph, > even for operators with a single input channel. > - End-users will see more progress even in unstable environments as more > up-to-date checkpoints will avoid too many recomputations. > - Facilitate faster rescaling. > > The key idea is to allow checkpoint barriers to be forwarded to downstream > tasks before the synchronous part of the checkpointing has been conducted > (see Fig. 1). To that end, we need to store in-flight data as part of the > checkpoint as described in greater details in this FLIP. > > Although the basic idea was already sketched in [2], we would like get > broader feedback in this dedicated mail thread. > > Best, > > Arvid > > [1] > https://cwiki.apache.org/confluence/display/FLINK/FLIP-76%3A+Unaligned+Checkpoints > [2] > http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/Checkpointing-under-backpressure-td31616.html >
