Yes, we are basically trying to implement the TS oracle talked about in the percolator paper. If we limit to a single host then that clearly has the downside of single point of failure, but is very easy to implement and is what we have now. I think we can do better and I want to lean on ZK for this because it seems like a good fit.
In terms of throughput, you can always batch up more request together if you need to. Basically in situations like this you can trade off latency for more throughput. One way to do this would be to have 2nd level timestamp servers that delegate to the leader TS server. These 2nd layer servers can smart batch up to 1K getFreshTs requests and send them all as one command to the leader TS server and it it can just do one AtomicLong.addAndGet(1K) and service all 1K of these in one request. That's what I was referring to before where the total latency will be O(lg(throughput)). You can extend this to 3 layers if you need to, but that is most likely not going to be needed. Getting millions/s will be pretty easy if we have 2 layers. Even with a single layer using smart batching I think we can get 1M/s ts handed out. Assuming no writes, we can get a ZK strongSync, checkExists done every 1ms or so and each of these can serve about 1K pending getFreshTs requests. That gets us to the 1M mark pretty easily without sacrificing the availability of running a single Ts node. This all assumes that there are enough transactions to go really wide. In the case where there is only one thread doing transactions, then throughput would be limited by the latency of the TS server which we'd like to bring down around 1-2ms if possible. -jc On Fri, Sep 28, 2012 at 8:59 AM, Flavio Junqueira <[email protected]> wrote: > I suppose you're aware of the Percolator approach to timestamps. According > to their description, they have a single server serving timestamps, which > makes performance very good in the absence of crashes of the timestamp > oracle. They mention restarts of the oracle, but they don't say how they > implement it; it is possibly done manually. > > Using master election to deal with oracle crashes requires that all > masters, primary and backups, agree on the sequence of timestamps handed to > clients. This observation essentially implies that you need go through zab > one way or the other to have agreement on the sequence of timestamps > handed. Even if we bypass disk writes but have it going though zab, I don't > think it will be able to get the same performance the timestamp oracle of > Percolator is getting (they claim 2 million/s). How much throughput > performance is sufficient for you if I may ask? Order of magnitude would be > sufficient. > > -Flavio > > On Sep 28, 2012, at 5:17 PM, John Carrino wrote: > > > Yes. This property needs to be global because otherwise another > transaction > > can start and commit in the past when it shouldn't be allowed to. > > > > -sent from my phone > > On Sep 28, 2012 8:13 AM, "Flavio Junqueira" <[email protected]> wrote: > > > >> I checked your message again and the contract is that > getFreshTimestamp() > >> always gets a fresh timestamp (not older than any returned by that > method). > >> One aspect of the contract that is not clear to me: does this freshness > >> property hold across clients or only for a single client? If it holds > only > >> per client, then you can guarantee it with master epochs. No user of > your > >> application will connect to the master of epoch y < x once it connects > to > >> the master of epoch x. The epochs can be the sequence number of a > >> sequential znode. How does it sound? > >> > >> -Flavio > >> > >> On Sep 28, 2012, at 5:00 PM, Flavio Junqueira wrote: > >> > >>> I was thinking that you can do a write per timestamp batch, and not per > >> individual timestamp. In the worst case, a former leader won't use some > >> timestamps, and I would think it is ok, but it depends on your > application. > >>> > >>> Also, even if two clients believe they are leaders simultaneously and > >> serve timestamps, the property that you seem to care about the most is > >> uniqueness: a timestamp is not served twice. Order of timestamps would > be > >> preserved per leader, but in the case of overlapping leaders you could > end > >> up serving timestamps that do not follow an increasing order. > >>> > >>> -Flavio > >>> > >>> On Sep 28, 2012, at 4:37 PM, John Carrino wrote: > >>> > >>>> CompareAndSwap is an atomic check and update. Basically only update > the > >>>> value if it is the same as the expected value. > >>>> > >>>> I think with your approach you'd have to do a write for every single > >>>> timestamp you wanted to hand out. The latency hit on this would too > >> much. > >>>> > >>>> My approach is different in that a timestamp server reserves a bunch > of > >>>> timestamps up front and proceeds to hand them out as long as it is the > >>>> leader. Leader check can be done without hitting disk hopefully. > >>>> > >>>> Thanks! > >>>> > >>>> -jc > >>>> > >>>> > >>>> On Fri, Sep 28, 2012 at 7:19 AM, Flavio Junqueira <[email protected]> > >> wrote: > >>>> > >>>>> I don't know what your compareAndSwap method does, but I was > wondering > >> if > >>>>> your client process can use conditional writes to a znode to make > sure > >> that > >>>>> it was the last one to update the state of timestamp batches. You can > >> treat > >>>>> the master election problem separately and it does not have to be as > >> strict > >>>>> as you have been thinking you need. Thats is, it wouldn't hurt if a > >> client > >>>>> still thinks it is leading even if it is not because no two clients > >> will be > >>>>> able to update the state of timestamp blocks without noticing that > >> another > >>>>> client is also updating it. > >>>>> > >>>>> -Flavio > >>>>> > >>>>> On Sep 27, 2012, at 6:57 PM, John Carrino wrote: > >>>>> > >>>>>> So I think it's time to explain what I'm writing just so everyone > has > >>>>>> more situation awareness. Its just a timestamp server, nothing > fancy. > >>>>>> > >>>>>> Looks like this: > >>>>>> > >>>>>> public interface TimestampService { > >>>>>> /** > >>>>>> * This will get a fresh timestamp that is guarenteed to be newer > >> than > >>>>>> any other timestamp > >>>>>> * handed out before this method was called. > >>>>>> */ > >>>>>> long getFreshTimestamp(); > >>>>>> } > >>>>>> > >>>>>> The only requirement is that the timestamp handed back is greater > than > >>>>>> every other timestamp that was returned before getFreshTs was > called. > >>>>>> There is no ordering requirement for concurrent requests. > >>>>>> > >>>>>> My impl is to reserve blocks of timestamps that are safe to hand out > >> (1M > >>>>> at > >>>>>> a time) using compare and swap in ZK. > >>>>>> lastPossibleUsed = read(HighWater) > >>>>>> safeToHandout = compareAndSwap(lastPossibleUsed, > lastPossibleUsed+1M) > >>>>>> > >>>>>> Now my leader can hand back timestamps up to safeToHandout, but > >> before it > >>>>>> hands one out it must ensure it is still the leader (no one else has > >>>>> handed > >>>>>> back something higher). > >>>>>> I can use ensureQuorum(), exists(myEphemNode) to make sure this is > the > >>>>>> case. Now I have a service that is guarenteed to be correct, but > >> doesn't > >>>>>> require disk hits in the steady state which brings down my latency > (if > >>>>> you > >>>>>> get close to running out, you can compareAndSwap for more > timestamps). > >>>>>> > >>>>>> If many requests come in at the same time I can use smart batching > to > >>>>>> verify happens after for all at once. We can also add more layers > if > >> we > >>>>>> need more bandwidth to scale up at the cost of adding latency. > >> Basically > >>>>>> our latency will be O(lg(requestRate)) if we keep adding layers as > >> each > >>>>>> previous layer becomes saturated. > >>>>>> > >>>>>> I hope this explanation helps. I am busy for the next 4 hours, but > if > >> you > >>>>>> need more clarification I can respond to them at that time. > >>>>>> > >>>>>> -jc > >>>>>> > >>>>>> > >>>>>> On Thu, Sep 27, 2012 at 9:26 AM, John Carrino < > [email protected] > >>>>>> wrote: > >>>>>> > >>>>>>> First, thanks everyone for talking this through with me. > >>>>>>> > >>>>>>> Flavio, for your example, this is actually ok. There is a happens > >> after > >>>>>>> relationship between the client making the request and my leader C1 > >>>>> still > >>>>>>> being the leader. My service only needs to guarantee that what it > >> hands > >>>>>>> back is at least as new as anything that existed when the client > made > >>>>> the > >>>>>>> request. If C2 were to answer requests while C1 is stalling that > is > >> ok > >>>>>>> because these would be considered concurrent requests and the stuff > >>>>>>> returned by C2 may be newer but that doesn't violate any > guarentees. > >>>>>>> > >>>>>>> If some client were to get back something from C2 and then (happens > >>>>> after > >>>>>>> relationship) someone tried to read from C1, it needs to fail. > >>>>>>> > >>>>>>> To address your concern of adding too much bandwidth we can get > this > >>>>>>> easily by doing what Martin Thompson calls smart batching ( > >>>>>>> > http://mechanical-sympathy.blogspot.com/2011/10/smart-batching.html > >> ). > >>>>>>> > >>>>>>> 1. ensureQuorum request comes in to L1 > >>>>>>> 2. send ENSURE to all followers > >>>>>>> 3. 10 more ensureQuorum requests come in > >>>>>>> 4. get back ENSURE from quorum > >>>>>>> 5. we can now service all 10 pending ensureQuorum requests with > >> another > >>>>>>> round trip ENSURE. > >>>>>>> > >>>>>>> We don't need to send an ENSURE for every ensureQuorum request, we > >> just > >>>>>>> need it to be happens after from when the request arrived. > >>>>>>> > >>>>>>> I am fine with the Ephemeral node being removed after some time > >> expires, > >>>>>>> but only by the leader. If the leaders clock is broken and the > >> client > >>>>>>> owning the Ephemeral node drops off, then we don't have liveness > >>>>> (because > >>>>>>> this node may not get cleaned up in a timely fashion). However, we > >>>>> still > >>>>>>> preserve corectness. > >>>>>>> > >>>>>>> -jc > >>>>>>> > >>>>>>> > >>>>>>> On Thu, Sep 27, 2012 at 9:02 AM, Flavio Junqueira < > [email protected] > >>>>>> wrote: > >>>>>>> > >>>>>>>> Say that we implement what you're suggesting. Could you check if > >> this > >>>>>>>> scenario can happen: > >>>>>>>> > >>>>>>>> 1- Client C1 is the current leader and it super boosted read to > make > >>>>> sure > >>>>>>>> it is still the leader; > >>>>>>>> 2- We process the super boosted read having it through the zab > >>>>> pipeline; > >>>>>>>> 3- When we send the response to C1 we slow down the whole deal: > the > >>>>>>>> response to C1 gets delayed and we stall C1; > >>>>>>>> 4- In the meanwhile, C1's session expires on the server side and > its > >>>>>>>> ephemeral leadership node is removed; > >>>>>>>> 5- A new client C2 is elected and starts exercising leadership; > >>>>>>>> 6- Now C1 comes back to normal and receives the response of the > >> super > >>>>>>>> boosted read saying that it is still the leader. > >>>>>>>> > >>>>>>>> If my interpretation is not incorrect, the only way to prevent > this > >>>>>>>> scenario from happening is if the session expires on the client > side > >>>>> before > >>>>>>>> it receives the response of the read. It doesn't look like we can > do > >>>>> it if > >>>>>>>> process clocks can be arbitrarily delayed. > >>>>>>>> > >>>>>>>> Note that one issue is that the behavior of ephemerals is highly > >>>>>>>> dependent upon timers, so I don't think we can avoid making some > >> timing > >>>>>>>> assumptions altogether. The question is if we are better off with > a > >>>>>>>> mechanism relying upon acknowledgements. My sense is that > >>>>> application-level > >>>>>>>> fencing is preferable (if not necessary) for applications like the > >>>>> ones JC > >>>>>>>> is mentioning or BookKeeper. > >>>>>>>> > >>>>>>>> I'm not concerned about writes to disk, which I agree we don't > need > >> for > >>>>>>>> sync. I'm more concerned about having it going through the whole > >>>>> pipeline, > >>>>>>>> which will induce more traffic to zab and increase latency for an > >>>>>>>> application that uses it heavily. > >>>>>>>> > >>>>>>>> -Flavio > >>>>>>>> > >>>>>>>> On Sep 27, 2012, at 5:27 PM, Alexander Shraer wrote: > >>>>>>>> > >>>>>>>>> another idea is to add this functionality to MultiOp - have read > >> only > >>>>>>>>> transactions be replicated but not logged or logged > asynchronously. > >>>>>>>>> I'm not sure how it works right now if I do a read-only MultiOp > >>>>>>>>> transaction - does it replicate the transaction or answer it > >> locally > >>>>>>>>> on the leader ? > >>>>>>>>> > >>>>>>>>> Alex > >>>>>>>>> > >>>>>>>>> On Thu, Sep 27, 2012 at 8:07 AM, Alexander Shraer < > >> [email protected]> > >>>>>>>> wrote: > >>>>>>>>>> Thanks for the explanation. > >>>>>>>>>> > >>>>>>>>>> I guess one could always invoke a write operation instead of > sync > >> to > >>>>>>>>>> get the more strict semantics, but as John suggests, it might > be a > >>>>>>>>>> good idea to add a new type of operation that requires followers > >> to > >>>>>>>>>> ack but doesn't require them to log to disk - this seems > >> sufficient > >>>>> in > >>>>>>>>>> our case. > >>>>>>>>>> > >>>>>>>>>> Alex > >>>>>>>>>> > >>>>>>>>>> On Thu, Sep 27, 2012 at 3:56 AM, Flavio Junqueira < > >> [email protected] > >>>>>> > >>>>>>>> wrote: > >>>>>>>>>>> In theory, the scenario you're describing could happen, but I > >> would > >>>>>>>> argue that it is unlikely given that: 1) a leader pings followers > >>>>> twice a > >>>>>>>> tick to make sure that it has a quorum of supporters (lead()); 2) > >>>>> followers > >>>>>>>> give up on a leader upon catching an exception (followLeader()). > One > >>>>> could > >>>>>>>> calibrate tickTime to make the probability of having this scenario > >> low. > >>>>>>>>>>> > >>>>>>>>>>> Let me also revisit the motivation for the way we designed > sync. > >>>>>>>> ZooKeeper has been designed to serve reads efficiently and making > >> sync > >>>>> go > >>>>>>>> through the pipeline would slow down reads. Although optional, we > >>>>> thought > >>>>>>>> it would be a good idea to make it as efficient as possible to > >> comply > >>>>> with > >>>>>>>> the original expectations for the service. We consequently came up > >> with > >>>>>>>> this cheap way of making sure that a read sees all pending > updates. > >> It > >>>>> is > >>>>>>>> correct that there are some corner cases that it doesn't cover. > One > >> is > >>>>> the > >>>>>>>> case you mentioned. Another is having the sync finishing before > the > >>>>> client > >>>>>>>> submits the read and having a write committing in between. We rely > >>>>> upon the > >>>>>>>> way we implement timeouts and some minimum degree of synchrony for > >> the > >>>>>>>> clients when submitting operations to guarantee that the scheme > >> work. > >>>>>>>>>>> > >>>>>>>>>>> We thought about the option of having the sync operation going > >>>>>>>> through the pipeline, and in fact it would have been easier to > >>>>> implement it > >>>>>>>> just as a regular write, but we opted not to because we felt it > was > >>>>>>>> sufficient for the use cases we had and more efficient as I > already > >>>>> argued. > >>>>>>>>>>> > >>>>>>>>>>> Hope it helps to clarify. > >>>>>>>>>>> > >>>>>>>>>>> -Flavio > >>>>>>>>>>> > >>>>>>>>>>> On Sep 27, 2012, at 9:38 AM, Alexander Shraer wrote: > >>>>>>>>>>> > >>>>>>>>>>>> thanks for the explanation! but how do you avoid having the > >>>>> scenario > >>>>>>>>>>>> raised by John ? > >>>>>>>>>>>> lets say you're a client connected to F, and F is connected to > >> L. > >>>>>>>> Lets > >>>>>>>>>>>> also say that L's pipeline > >>>>>>>>>>>> is now empty, and both F and L are partitioned from 3 other > >> servers > >>>>>>>> in > >>>>>>>>>>>> the system that have already > >>>>>>>>>>>> elected a new leader L'. Now I go to L' and write something. L > >>>>> still > >>>>>>>>>>>> thinks its the leader because the > >>>>>>>>>>>> detection that followers left it is obviously timeout > >> dependent. So > >>>>>>>>>>>> when F sends your sync to L and L returns > >>>>>>>>>>>> it to F, you actually miss my write! > >>>>>>>>>>>> > >>>>>>>>>>>> Alex > >>>>>>>>>>>> > >>>>>>>>>>>> On Thu, Sep 27, 2012 at 12:32 AM, Flavio Junqueira < > >>>>>>>> [email protected]> wrote: > >>>>>>>>>>>>> Hi Alex, Because of the following: > >>>>>>>>>>>>> > >>>>>>>>>>>>> 1- A follower F processes operations from a client in FIFO > >> order, > >>>>>>>> and say that a client submits as you say sync + read; > >>>>>>>>>>>>> 2- A sync will be processed by the leader and returned to the > >>>>>>>> follower. It will be queued after all pending updates that the > >> follower > >>>>>>>> hasn't processed; > >>>>>>>>>>>>> 3- The follower will process all pending updates before > >> processing > >>>>>>>> the response of the sync; > >>>>>>>>>>>>> 4- Once the follower processes the sync, it picks the read > >>>>>>>> operation to process. It reads the local state of the follower and > >>>>> returns > >>>>>>>> to the client. > >>>>>>>>>>>>> > >>>>>>>>>>>>> When we process the read in Step 4, we have applied all > pending > >>>>>>>> updates the leader had for the follower by the time the read > request > >>>>>>>> started. > >>>>>>>>>>>>> > >>>>>>>>>>>>> This implementation is a bit of a hack because it doesn't > >> follow > >>>>>>>> the same code path as the other operations that go to the leader, > >> but > >>>>> it > >>>>>>>> avoids some unnecessary steps, which is important for fast reads. > In > >>>>> the > >>>>>>>> sync case, the other followers don't really need to know about it > >>>>> (there is > >>>>>>>> nothing to be updated) and the leader simply inserts it in the > >>>>> sequence of > >>>>>>>> updates of F, ordering it. > >>>>>>>>>>>>> > >>>>>>>>>>>>> -Flavio > >>>>>>>>>>>>> > >>>>>>>>>>>>> On Sep 27, 2012, at 9:12 AM, Alexander Shraer wrote: > >>>>>>>>>>>>> > >>>>>>>>>>>>>> Hi Flavio, > >>>>>>>>>>>>>> > >>>>>>>>>>>>>>> Starting a read operation concurrently with a sync implies > >> that > >>>>>>>> the result of the read will not miss an update committed before > the > >>>>> read > >>>>>>>> started. > >>>>>>>>>>>>>> > >>>>>>>>>>>>>> I thought that the intention of sync was to give something > >> like > >>>>>>>>>>>>>> linearizable reads, so if you invoke a sync and then a read, > >> your > >>>>>>>> read > >>>>>>>>>>>>>> is guaranteed to (at least) see any write which completed > >> before > >>>>>>>> the > >>>>>>>>>>>>>> sync began. Is this the intention ? If so, how is this > >> achieved > >>>>>>>>>>>>>> without running agreement on the sync op ? > >>>>>>>>>>>>>> > >>>>>>>>>>>>>> Thanks, > >>>>>>>>>>>>>> Alex > >>>>>>>>>>>>>> > >>>>>>>>>>>>>> On Thu, Sep 27, 2012 at 12:05 AM, Flavio Junqueira < > >>>>>>>> [email protected]> wrote: > >>>>>>>>>>>>>>> sync simply flushes the channel between the leader and the > >>>>>>>> follower that forwarded the sync operation, so it doesn't go > through > >>>>> the > >>>>>>>> full zab pipeline. Flushing means that all pending updates from > the > >>>>> leader > >>>>>>>> to the follower are received by the time sync completes. Starting > a > >>>>> read > >>>>>>>> operation concurrently with a sync implies that the result of the > >> read > >>>>> will > >>>>>>>> not miss an update committed before the read started. > >>>>>>>>>>>>>>> > >>>>>>>>>>>>>>> -Flavio > >>>>>>>>>>>>>>> > >>>>>>>>>>>>>>> On Sep 27, 2012, at 3:43 AM, Alexander Shraer wrote: > >>>>>>>>>>>>>>> > >>>>>>>>>>>>>>>> Its strange that sync doesn't run through agreement, I was > >>>>> always > >>>>>>>>>>>>>>>> assuming that it is... Exactly for the reason you say - > >>>>>>>>>>>>>>>> you may trust your leader, but I may have a different > leader > >>>>> and > >>>>>>>> your > >>>>>>>>>>>>>>>> leader may not detect it yet and still think its the > leader. > >>>>>>>>>>>>>>>> > >>>>>>>>>>>>>>>> This seems like a bug to me. > >>>>>>>>>>>>>>>> > >>>>>>>>>>>>>>>> Similarly to Paxos, Zookeeper's safety guarantees don't > (or > >>>>>>>> shouldn't) > >>>>>>>>>>>>>>>> depend on timing assumption. > >>>>>>>>>>>>>>>> Only progress guarantees depend on time. > >>>>>>>>>>>>>>>> > >>>>>>>>>>>>>>>> Alex > >>>>>>>>>>>>>>>> > >>>>>>>>>>>>>>>> > >>>>>>>>>>>>>>>> On Wed, Sep 26, 2012 at 4:41 PM, John Carrino < > >>>>>>>> [email protected]> wrote: > >>>>>>>>>>>>>>>>> I have some pretty strong requirements in terms of > >> consistency > >>>>>>>> where > >>>>>>>>>>>>>>>>> reading from followers that may be behind in terms of > >> updates > >>>>>>>> isn't ok for > >>>>>>>>>>>>>>>>> my use case. > >>>>>>>>>>>>>>>>> > >>>>>>>>>>>>>>>>> One error case that worries me is if a follower and > leader > >> are > >>>>>>>> partitioned > >>>>>>>>>>>>>>>>> off from the network. A new leader is elected, but the > >>>>>>>> follower and old > >>>>>>>>>>>>>>>>> leader don't know about it. > >>>>>>>>>>>>>>>>> > >>>>>>>>>>>>>>>>> Normally I think sync was made for this purpost, but I > >> looked > >>>>>>>> at the sync > >>>>>>>>>>>>>>>>> code and if there aren't any outstanding proposals the > >> leader > >>>>>>>> sends the > >>>>>>>>>>>>>>>>> sync right back to the client without first verifying > that > >> it > >>>>>>>> still has > >>>>>>>>>>>>>>>>> quorum, so this won't work for my use case. > >>>>>>>>>>>>>>>>> > >>>>>>>>>>>>>>>>> At the core of the issue all I really need is a call that > >> will > >>>>>>>> make it's > >>>>>>>>>>>>>>>>> way to the leader and will ping it's followers, ensure it > >>>>> still > >>>>>>>> has a > >>>>>>>>>>>>>>>>> quorum and return success. > >>>>>>>>>>>>>>>>> > >>>>>>>>>>>>>>>>> Basically a getCurrentLeaderEpoch() method that will be > >>>>>>>> forwarded to the > >>>>>>>>>>>>>>>>> leader, leader will ensure it still has quorum and return > >> it's > >>>>>>>> epoch. I > >>>>>>>>>>>>>>>>> can use this primitive to implement all the other > >> properties I > >>>>>>>> want to > >>>>>>>>>>>>>>>>> verify (assuming that my client will never connect to an > >> older > >>>>>>>> epoch after > >>>>>>>>>>>>>>>>> this call returns). Also the nice thing about this method > >> is > >>>>>>>> that it will > >>>>>>>>>>>>>>>>> not have to hit disk and the latency should just be a > round > >>>>>>>> trip to the > >>>>>>>>>>>>>>>>> followers. > >>>>>>>>>>>>>>>>> > >>>>>>>>>>>>>>>>> Most of the guarentees offered by zookeeper are time > based > >> an > >>>>>>>> rely on > >>>>>>>>>>>>>>>>> clocks and expiring timers, but I'm hoping to offer some > >>>>>>>> guarantees in > >>>>>>>>>>>>>>>>> spite of busted clocks, horrible GC perf, VM suspends and > >> any > >>>>>>>> other way > >>>>>>>>>>>>>>>>> time is broken. > >>>>>>>>>>>>>>>>> > >>>>>>>>>>>>>>>>> Also if people are interested I can go into more detail > >> about > >>>>>>>> what I am > >>>>>>>>>>>>>>>>> trying to write. > >>>>>>>>>>>>>>>>> > >>>>>>>>>>>>>>>>> -jc > >>>>>>>>>>>>>>> > >>>>>>>>>>>>> > >>>>>>>>>>> > >>>>>>>> > >>>>>>>> > >>>>>>> > >>>>> > >>>>> > >>> > >> > >> > >
