(1) the prepare marker is written, but the endTxn response is not received by the client when the server downgrades (2) the prepare marker is written, the endTxn response is received by the client when the server downgrades.
I think I am still a little confused. In both of these cases, the transaction log has the old producer ID. We don't write the new producer ID in the prepare marker's non tagged fields. If the server downgrades now, it would read the records not in tagged fields and the complete marker will also have the old producer ID. (If we had used the new producer ID, we would not have transactional correctness since the producer id doesn't match the transaction and the state would not be correct on the data partition.) In the overflow case, I'd expect the following to happen on the client side Case 1 -- we retry EndTxn -- it is the same producer ID and epoch - 1 this would fence the producer Case 2 -- we don't retry EndTxn and use the new producer id which would result in InvalidPidMappingException Maybe we can have special handling for when a server downgrades. When it reconnects we could get an API version request showing KIP-890 part 2 is not supported. In that case, we can call initProducerId to abort the transaction. (In the overflow case, this correctly gives us a new producer ID) I guess the corresponding case would be where the *complete marker *is written but the endTxn is not received by the client and the server downgrades? This would result in the transaction coordinator having the new ID and not the old one. If the client retries, it will receive an InvalidPidMappingException. The InitProducerId scenario above would help here too. To be clear, my compatibility story is meant to support downgrades server side in keeping the transactional correctness. Keeping the client from fencing itself is not the priority. Hope this helps. I can also add text in the KIP about InitProducerId if we think that fixes some edge cases. Justine On Fri, Jan 12, 2024 at 4:10 PM Jun Rao <j...@confluent.io.invalid> wrote: > Hi, Justine, > > Thanks for the reply. > > I agree that we don't need to optimize for fencing during downgrades. > Regarding consistency, there are two possible cases: (1) the prepare marker > is written, but the endTxn response is not received by the client when the > server downgrades; (2) the prepare marker is written, the endTxn response > is received by the client when the server downgrades. In (1), the client > will have the old produce Id and in (2), the client will have the new > produce Id. If we downgrade right after the prepare marker, we can't be > consistent to both (1) and (2) since we can only put one value in the > existing produce Id field. It's also not clear which case is more likely. > So we could probably be consistent with either case. By putting the new > producer Id in the prepare marker, we are consistent with case (2) and it > also has the slight benefit that the produce field in the prepare and > complete marker are consistent in the overflow case. > > Jun > > On Fri, Jan 12, 2024 at 3:11 PM Justine Olshan > <jols...@confluent.io.invalid> > wrote: > > > Hi Jun, > > > > In the case you describe, we would need to have a delayed request, send a > > successful EndTxn, and a successful AddPartitionsToTxn and then have the > > delayed EndTxn request go through for a given producer. > > I'm trying to figure out if it is possible for the client to transition > if > > a previous request is delayed somewhere. But yes, in this case I think we > > would fence the client. > > > > Not for the overflow case. In the overflow case, the producer ID and the > > epoch are different on the marker and on the new transaction. So we want > > the marker to use the max epoch but the new transaction should start > with > > the new ID and epoch 0 in the transactional state. > > > > In the server downgrade case, we want to see the producer ID as that is > > what the client will have. If we complete the commit, and the transaction > > state is reloaded, we need the new producer ID in the state so there > isn't > > an invalid producer ID mapping. > > The server downgrade cases are considering transactional correctness and > > not regressing from previous behavior -- and are not concerned about > > supporting the safety from fencing retries (as we have downgraded so we > > don't need to support). Perhaps this is a trade off, but I think it is > the > > right one. > > > > (If the client downgrades, it will have restarted and it is ok for it to > > have a new producer ID too). > > > > Justine > > > > On Fri, Jan 12, 2024 at 11:42 AM Jun Rao <j...@confluent.io.invalid> > wrote: > > > > > Hi, Justine, > > > > > > Thanks for the reply. > > > > > > 101.4 "If the marker is written by the new client, we can as I > mentioned > > in > > > the last email guarantee that any EndTxn requests with the same epoch > are > > > from the same producer and the same transaction. Then we don't have to > > > return a fenced error but can handle gracefully as described in the > KIP." > > > When a delayed EndTnx request is processed, the txn state could be > > ongoing > > > for the next txn. I guess in this case we still return the fenced error > > for > > > the delayed request? > > > > > > 102. Sorry, my question was inaccurate. What you described is accurate. > > > "The downgrade compatibility I mention is that we keep the same > producer > > ID > > > and epoch in the main (non-tagged) fields as we did before the code on > > the > > > server side." If we want to do this, it seems that we should use the > > > current produce Id and max epoch in the existing producerId and > > > producerEpoch fields for both the prepare and the complete marker, > right? > > > The downgrade can happen after the complete marker is written. With > what > > > you described, the downgraded coordinator will see the new produce Id > > > instead of the old one. > > > > > > Jun > > > > > > On Fri, Jan 12, 2024 at 10:44 AM Justine Olshan > > > <jols...@confluent.io.invalid> wrote: > > > > > > > Hi Jun, > > > > > > > > I can update the description. > > > > > > > > I believe your second point is mentioned in the KIP. I can add more > > text > > > on > > > > this if it is helpful. > > > > > The delayed message case can also violate EOS if the delayed > message > > > > comes in after the next addPartitionsToTxn request comes in. > > Effectively > > > we > > > > may see a message from a previous (aborted) transaction become part > of > > > the > > > > next transaction. > > > > > > > > If the marker is written by the new client, we can as I mentioned in > > the > > > > last email guarantee that any EndTxn requests with the same epoch are > > > from > > > > the same producer and the same transaction. Then we don't have to > > return > > > a > > > > fenced error but can handle gracefully as described in the KIP. > > > > I don't think a boolean is useful since it is directly encoded by the > > > > existence or lack of the tagged field being written. > > > > In the prepare marker we will have the same producer ID in the > > non-tagged > > > > field. In the Complete state we may not. > > > > I'm not sure why the ongoing state matters for this KIP. It does > matter > > > for > > > > KIP-939. > > > > > > > > I'm not sure what you are referring to about writing the previous > > > producer > > > > ID in the prepare marker. This is not in the KIP. > > > > In the overflow case, we write the nextProducerId in the prepare > state. > > > > This is so we know what we assigned when we reload the transaction > log. > > > > Once we complete, we transition this ID to the main (non-tagged > field) > > > and > > > > have the previous producer ID field filled in. This is so we can > > identify > > > > in a retry case the operation completed successfully and we don't > fence > > > our > > > > producer. The downgrade compatibility I mention is that we keep the > > same > > > > producer ID and epoch in the main (non-tagged) fields as we did > before > > > the > > > > code on the server side. If the server downgrades, we are still > > > compatible. > > > > This addresses both the prepare and complete state downgrades. > > > > > > > > Justine > > > > > > > > On Fri, Jan 12, 2024 at 10:21 AM Jun Rao <j...@confluent.io.invalid> > > > wrote: > > > > > > > > > Hi, Justine, > > > > > > > > > > Thanks for the reply. Sorry for the delay. I have a few more > > comments. > > > > > > > > > > 110. I think the motivation section could be improved. One of the > > > > > motivations listed by the KIP is "This can happen when a message > gets > > > > stuck > > > > > or delayed due to networking issues or a network partition, the > > > > transaction > > > > > aborts, and then the delayed message finally comes in.". This seems > > not > > > > > very accurate. Without KIP-890, currently, if the coordinator times > > out > > > > and > > > > > aborts an ongoing transaction, it already bumps up the epoch in the > > > > marker, > > > > > which prevents the delayed produce message from being added to the > > user > > > > > partition. What can cause a hanging transaction is that the > producer > > > > > completes (either aborts or commits) a transaction before > receiving a > > > > > successful ack on messages published in the same txn. In this case, > > > it's > > > > > possible for the delayed message to be appended to the partition > > after > > > > the > > > > > marker, causing a transaction to hang. > > > > > > > > > > A similar issue (not mentioned in the motivation) could happen on > the > > > > > marker in the coordinator's log. For example, it's possible for an > > > > > EndTxnRequest to be delayed on the coordinator. By the time the > > delayed > > > > > EndTxnRequest is processed, it's possible that the previous txn has > > > > already > > > > > completed and a new txn has started. Currently, since the epoch is > > not > > > > > bumped on every txn, the delayed EndTxnRequest will add an > unexpected > > > > > prepare marker (and eventually a complete marker) to the ongoing > txn. > > > > This > > > > > won't cause the transaction to hang, but it will break the EoS > > > semantic. > > > > > The proposal in this KIP will address this issue too. > > > > > > > > > > 101. "However, I was writing it so that we can distinguish between > > > > > old clients where we don't have the ability do this operation and > new > > > > > clients that can. (Old clients don't bump the epoch on commit, so > we > > > > can't > > > > > say for sure the write belongs to the given transaction)." > > > > > 101.1 I am wondering why we need to distinguish whether the marker > is > > > > > written by the old and the new client. Could you describe what we > do > > > > > differently if we know the marker is written by the new client? > > > > > 101.2 If we do need a way to distinguish whether the marker is > > written > > > by > > > > > the old and the new client. Would it be simpler to just introduce a > > > > boolean > > > > > field instead of indirectly through the previous produce ID field? > > > > > 101.3 It's not clear to me why we only add the previous produce ID > > > field > > > > in > > > > > the complete marker, but not in the prepare marker. If we want to > > know > > > > > whether a marker is written by the new client or not, it seems that > > we > > > > want > > > > > to do this consistently for all markers. > > > > > 101.4 What about the TransactionLogValue record representing the > > > ongoing > > > > > state? Should we also distinguish whether it's written by the old > or > > > the > > > > > new client? > > > > > > > > > > 102. In the overflow case, it's still not clear to me why we write > > the > > > > > previous produce Id in the prepare marker while writing the next > > > produce > > > > Id > > > > > in the complete marker. You mentioned that it's for downgrading. > > > However, > > > > > we could downgrade with either the prepare marker or the complete > > > marker. > > > > > In either case, the downgraded coordinator should see the same > > produce > > > id > > > > > (probably the previous produce Id), right? > > > > > > > > > > Jun > > > > > > > > > > On Wed, Dec 20, 2023 at 6:00 PM Justine Olshan > > > > > <jols...@confluent.io.invalid> > > > > > wrote: > > > > > > > > > > > Hey Jun, > > > > > > > > > > > > Thanks for taking a look at the KIP again. > > > > > > > > > > > > 100. For the epoch overflow case, only the marker will have max > > > epoch. > > > > > This > > > > > > keeps the behavior of the rest of the markers where the last > marker > > > is > > > > > the > > > > > > epoch of the transaction records + 1. > > > > > > > > > > > > 101. You are correct that we don't need to write the producer ID > > > since > > > > it > > > > > > is the same. However, I was writing it so that we can distinguish > > > > between > > > > > > old clients where we don't have the ability do this operation and > > new > > > > > > clients that can. (Old clients don't bump the epoch on commit, so > > we > > > > > can't > > > > > > say for sure the write belongs to the given transaction). If we > > > receive > > > > > an > > > > > > EndTxn request from a new client, we will fill this field. We can > > > > > guarantee > > > > > > that any EndTxn requests with the same epoch are from the same > > > producer > > > > > and > > > > > > the same transaction. > > > > > > > > > > > > 102. In prepare phase, we have the same producer ID and epoch we > > > always > > > > > > had. It is the producer ID and epoch that are on the marker. In > > > commit > > > > > > phase, we stay the same unless it is the overflow case. In that > > case, > > > > we > > > > > > set the producer ID to the new one we generated and epoch to 0 > > after > > > > > > complete. This is for downgrade compatibility. The tagged fields > > are > > > > just > > > > > > safety guards for retries and failovers. > > > > > > > > > > > > In prepare phase for epoch overflow case only we store the next > > > > producer > > > > > > ID. This is for the case where we reload the transaction > > coordinator > > > in > > > > > > prepare state. Once the transaction is committed, we can use the > > > > producer > > > > > > ID the client already is using. > > > > > > > > > > > > In commit phase, we store the previous producer ID in case of > > > retries. > > > > > > > > > > > > I think it is easier to think of it as just how we were storing > > > > producer > > > > > ID > > > > > > and epoch before, with some extra bookeeping and edge case > handling > > > in > > > > > the > > > > > > tagged fields. We have to do it this way for compatibility with > > > > > downgrades. > > > > > > > > > > > > 103. Next producer ID is for prepare status and previous producer > > ID > > > is > > > > > for > > > > > > after complete. The reason why we need two separate (tagged) > fields > > > is > > > > > for > > > > > > backwards compatibility. We need to keep the same semantics for > the > > > > > > non-tagged field in case we downgrade. > > > > > > > > > > > > 104. We set the fields as we do in the transactional state (as we > > > need > > > > to > > > > > > do this for compatibility -- if we downgrade, we will only have > the > > > > > > non-tagged fields) It will be the old producer ID and max epoch. > > > > > > > > > > > > Hope this helps. Let me know if you have further questions. > > > > > > > > > > > > Justine > > > > > > > > > > > > On Wed, Dec 20, 2023 at 3:33 PM Jun Rao <j...@confluent.io.invalid > > > > > > > wrote: > > > > > > > > > > > > > Hi, Justine, > > > > > > > > > > > > > > It seems that you have made some changes to KIP-890 since the > > vote. > > > > In > > > > > > > particular, we are changing the format of TransactionLogValue. > A > > > few > > > > > > > comments related to that. > > > > > > > > > > > > > > 100. Just to be clear. The overflow case (i.e. when a new > > > producerId > > > > is > > > > > > > generated) is when the current epoch equals to max - 1 and not > > max? > > > > > > > > > > > > > > 101. For the "not epoch overflow" case, we write the previous > ID > > in > > > > the > > > > > > > tagged field in the complete phase. Do we need to do that since > > > > produce > > > > > > id > > > > > > > doesn't change in this case? > > > > > > > > > > > > > > 102. It seems that the meaning for the ProducerId/ProducerEpoch > > > > fields > > > > > in > > > > > > > TransactionLogValue changes depending on the TransactionStatus. > > > When > > > > > > > the TransactionStatus is ongoing, they represent the current > > > > ProducerId > > > > > > and > > > > > > > the current ProducerEpoch. When the TransactionStatus is > > > > > > > PrepareCommit/PrepareAbort, they represent the current > ProducerId > > > and > > > > > the > > > > > > > next ProducerEpoch. When the TransactionStatus is Commit/Abort, > > > they > > > > > > > further depend on whether the epoch overflows or not. If there > is > > > no > > > > > > > overflow, they represent the current ProducerId and the next > > > > > > ProducerEpoch > > > > > > > (max). Otherwise, they represent the newly generated ProducerId > > > and a > > > > > > > ProducerEpoch of 0. Is that right? This seems not easy to > > > understand. > > > > > > Could > > > > > > > we provide some examples like what Artem has done in KIP-939? > > Have > > > we > > > > > > > considered a simpler design where ProducerId/ProducerEpoch > always > > > > > > represent > > > > > > > the same value (e.g. for the current transaction) independent > of > > > the > > > > > > > TransactionStatus and epoch overflow? > > > > > > > > > > > > > > 103. It's not clear to me why we need 3 fields: ProducerId, > > > > > > PrevProducerId, > > > > > > > NextProducerId. Could we just have ProducerId and > NextProducerId? > > > > > > > > > > > > > > 104. For WriteTxnMarkerRequests, if the producer epoch > overflows, > > > > what > > > > > do > > > > > > > we set the producerId and the producerEpoch? > > > > > > > > > > > > > > Thanks, > > > > > > > > > > > > > > Jun > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >
