Hi Hao, Thanks to your reply, hao.
you are proposing to use key->ops hash to apply the write transaction instead of serializing the operations in PREPARE phrase with a lock? Changing thread mode from concurrent to serial in APPLY phase. Splitting TXN to ops, enqueuing the op to apply_pool token based on hash(primary-key), then the op(primary key) can be applied in turn. 1) what the time stamp would be if there is no longer a lock during PREPARE phrase, how to ensure the following scenario happens in order? * T1 deletes the row (key1, key2, value1) * T2 upserts the row (key1, key2, value2) It is ordered in PREPARE phase, so timestamp will be increased as TXN order, and also op-index, raft protocol guarantee TXN are orderly submitted to apply_pool_ by op-index. We just keep keys are ordered to apply in APPLY phase after abandoning row-key lock, as the plan says. 2) what happens if any of the operations happens to update the keys? e.g * T1 updates the row (key1, key2, value1) to (key3, key2, value2) * T1 updates the row (key3, key2, value3) to (key1, key2, value1) Like to be described above, it is ordered to apply by key. I will submit a JIRA and attach the design doc later. Hope and please you keep a eye about that. Thanks. ----- Regards, Xiaokai ________________________________ From: Hao Hao <[email protected]> Sent: Monday, September 24, 2018 11:18 PM To: [email protected] Subject: Re: Locks are acquired to cost much time in transactions Hi Xiaokai, If I understand you correctly, you are proposing to use key->ops hash to apply the write transaction instead of serializing the operations in PREPARE phrase with a lock? Here are some questions I have with this approach, 1) what the time stamp would be if there is no longer a lock during PREPARE phrase, how to ensure the following scenario happens in order? * T1 deletes the row (key1, key2, value1) * T2 upserts the row (key1, key2, value2) 2) what happens if any of the operations happens to update the keys? e.g * T1 updates the row (key1, key2, value1) to (key3, key2, value2) * T1 updates the row (key3, key2, value3) to (key1, key2, value1) Both of the above scenarios don't seem to have a deterministic behavior. Best, Hao On Fri, Sep 21, 2018 at 2:29 AM, Xiaokai Wang <[email protected]<mailto:[email protected]>> wrote: Thanks your reply, Mike. In our scenario, some tables have a lot of data to be updated at the same time. The current implementation of KUDU, if TXN(transaction) can not get all the ops lock it will wait for 1s to block the entire Tablet processing, which will greatly reduce the write performance of the Tablet.Reduce the size of the batch, TXN may still have the same row key, and can not solve this problem. In my way, I want to change the work way of 'APPLY' phase, instead of putting TXN to apply_pool_ queue, I will put each key-op(splitting TXN to ops) to apply_pool_token_ queue. This will guarantee the same key's op will be putted to the same thread. When ops belonging to the same TXN execute over, then send rpc response and write CommitMsg to WAL. In this way, I can abandon the keys locks. Working follow chart just like this below: [op-key_concurrent.png] 1. client send T1, T2, T3 to tserver, tserver will handle each txn in turn. 2. leader send replicated msg to follower, sending them together or independent. T1, T2, T3 will be ordered to execute by follower and send them back by together or independent in turn. raft_pool_token_ guarantee T1, T2 T3 be ordered to apply. 3. Splitting Tx to ops, each op is hashed to the queue of apply_pool_token_, key's op is order to execute as TXN turn. -------- Kudu origin work way below: [txn_concurrent.png] 1. client send T1, T2, T3 to tserver, tserver will handle each txn orderly, they will acquire locks in 'PREPARE' phase in turn. 2. leader send replicated msg to follower, sending them together or independent. T1, T2, T3 will be ordered to execute by follower and send them back by together or independent in turn. raft_pool_token_ guarantee T1, T2 T3 be ordered to apply. 3. T1, T2, T3 are putted to the queue of apply_pool_, order or concurrented to be executed, releasing locks. ----------- Contrast to origin, the keys locks are abandon, this can guarantee kudu throughput more smoothly. What do you think? Hope to get your advice, kudu users. Thanks. xiaokai. ---------- Forwarded message --------- From: Mike Percy <[email protected]<mailto:[email protected]>> Date: 2018年9月19日周三 上午8:29 Subject: Re: Locks are acquired to cost much time in transactions To: <[email protected]<mailto:[email protected]>> Why do you think you are spending a lot of time contending on row locks? Have you tried configuring your clients to send smaller batches? This may decrease throughput on a per-client basis but will likely improve latency and reduce the likelihood of row lock contention. If you are really spending most of your time contending on row locks then you will likely run into more fundamental performance issues trying to scale your writes, since Kudu's MVCC implementation effectively stores a linked list of updates to a given cell until compaction occurs. See https://github.com/apache/kudu/blob/master/docs/design-docs/tablet.md#historical-mvcc-in-diskrowsets for more information about the on-disk design. If you accumulate too many uncompacted mutations against a given row, reading the latest value for that row at scan time will be slow because it has to do a lot of work at read time. Mike On Tue, Sep 18, 2018 at 8:48 AM Xiaokai Wang <[email protected]<mailto:[email protected]>> wrote: Moved here from JIRA. Hi guys, I met a problem about the keys locks that almost impacts the service normal writing. As we all know, a transaction which get all row_key locks will go on next step in kudu. Everything looks good, if keys are not concurrent updated. But when keys are updated by more than one client at the same time, locks are acquired to wait much time. The cases are often in my product environment. Does anybody meet the problem? Has any good ideal for this? In my way, I want to try to abandon keys locks, instead using *_pool_token_ 'SERIAL' mode which keeping the key of transaction is serial and ordered. Dose this work? Hope to get your advice. Thanks. ----- Regards, Xiaokai
