Igniters,
I've looked through Nikita's changes and I think for the current issue [1] we should not allow the existence of MOVING partitions in the cluster (it must be stable) to run the lightweight PME on BLT node leave event occurred to achieve truly unlocked operations and here are my thoughts why. In general, as Nikita mentioned above, the existence of MOVING partitions in the cluster means that the rebalance procedure is currently running. It owns cache partitions locally and sends in the background (with additional timeout) the actual statuses of his local partitions to the coordinator node. So, we will always have a lag between local node partition states and all other cluster nodes partitions states. This lag can be very huge since previous #scheduleResendPartitions() is cancelled when a new cache group rebalance finished. Without the fair partition states synchronization (without full PME) and in case of local affinity recalculation on BLT node leave event, other nodes will mark such partitions LOST in most of the cases, which in fact are present in the cluster and saved on some node under checkpoint. I see that it cannot be solved by saving transition states of such partitions on each node. As for the case when the coordinator will calculate affinity and send "full map" to other nodes, I think it is better here to focus on designing a new lightweight PME when the rebalancing process finishes. Сurrently full distributed PME will occur anyway by the coordinator by sending CacheAffinityChaneMessage, but I think we can avoid it here, since no new MOVING or OWNING node partition states are introduced and all the previous mappings are still valid. We don't need a distributed PME if we will leave partition primaries on those nodes where they were, just set correct partition statuses via a light discovery message. So, my plan here can be: Phase 1. Lightweight PME on BLT node leave on a stable cluster (no MOVING partitions); Phase 2. Lightweight PME on BLT node finishes its rebalance procedure. Folks, Nikita, WDYT? [1] https://issues.apache.org/jira/browse/IGNITE-9913 On Fri, 24 May 2019 at 13:31, Nikita Amelchev <nsamelc...@gmail.com> wrote: > > Hello, Igniters! > > I am working on the implementation of lightweight PME for the case of > a BLT node leave. [1] > > There is a question: whether to allow lightweight PME if the cluster > has MOVING partitions? > > The problems that may happen if allow: > - Nodes can differently select the primary node from current OWNING backups. > - One part of nodes can mark a partition as LOST and another one as OWNING. > > We can take states of the partitions from the node2part map. The root > cause of those problems is that when rebalancing ends (get the last > message), it updates partition state of the local node to OWNING (and > schedules partitions resend). This may lead to different affinity > re-calculations on nodes. > > I see two solutions: > > 1. Nodes will store “moving-owning” transition of partitions state > until the rebalancing ends. Each node will locally recalculate the > affinity on node left event. > 2. The coordinator will calculate affinity and send "full map" to > nodes. In this case, nodes still should wait for topology change event > (to get correct topology in discovery). > > If disallow lightweight PME when the cluster has MOVING partitions - > there are no problems and it works fine. > > Any thoughts? > > 1. https://issues.apache.org/jira/browse/IGNITE-9913 > > пт, 29 мар. 2019 г. в 15:00, Nikita Amelchev <nsamelc...@gmail.com>: > > > > Pavel, > > I have provided MTCGA bot status in Jira issue comments. [1] > > > > Eduard, > > Yes, for current implementation it will be distributed PME if > > in-memory caches configured. > > > > 1. https://issues.apache.org/jira/browse/IGNITE-9913 > > > > пт, 29 мар. 2019 г. в 14:49, Eduard Shangareev > > <eduard.shangar...@gmail.com>: > > > > > > Nikita, > > > > > > It sounds cool. But I didn't get about in-memory caches. The baseline is > > > not used for their affinity calculation. > > > So, this improvement would be switched off for them or completely (when > > > such caches are presented), wouldn't it? > > > > > > On Thu, Mar 28, 2019 at 3:14 PM Pavel Kovalenko <jokse...@gmail.com> > > > wrote: > > > > > > > Hi Nikita, > > > > > > > > Thank you for your work. This is great improvement. I'll take look on > > > > it in > > > > next couple of days. Could you please run TC and provide MTCGA bot > > > > status > > > > about this change? > > > > > > > > чт, 28 мар. 2019 г. в 14:29, Nikita Amelchev <nsamelc...@gmail.com>: > > > > > > > > > Hello, Igniters! > > > > > > > > > > I have implemented lightweight version of partitions map exchange for > > > > > the case when the baseline node leaves topology. [1] > > > > > > > > > > If partitions are assigned according to the baseline topology and > > > > > server node leaves there's no actual need to perform distributed PME. > > > > > Every cluster will recalculate new affinity assignments and partition > > > > > states locally. There is no need to wait for partitions released and > > > > > PME will be started immediately. > > > > > > > > > > I have benchmarked duration of PME under yardstick load. PME duration > > > > > was decreased up to 10 times and the maximum latency of transactions > > > > > was decreased up to 4-5 times. See details in Jira issue comments. [1] > > > > > > > > > > Could some expert of PME take a look at my changes? [2] > > > > > > > > > > 1. https://issues.apache.org/jira/browse/IGNITE-9913 > > > > > 2. https://reviews.ignite.apache.org/ignite/review/IGNT-CR-1027 > > > > > > > > > > -- > > > > > Best wishes, > > > > > Amelchev Nikita > > > > > > > > > > > > > > > > > -- > > Best wishes, > > Amelchev Nikita > > > > -- > Best wishes, > Amelchev Nikita
