Hi Dmitriy, It is a pitty that we have to guess here. Some comments: 1. The usage scenario is not very complex and erything might work well with aforementioned patch. Unfortunately some tradeoff and extra transformations are involved and consequently it is hard to predict how many users will be troubled. But on a bright side the change is seemless and perhaps nobody will be really troubled. 2. From "honest" code perspective introducing nodeStorageName seems better to me because no (surprising) transformations are involved. 3. Also, it seems to me that original logic for choosing storage was not designed well. From the first glance it looks strange that need folders for a particular node in different places. Is not it much clear to have all data of the node in a single directory isolated from other nodes? AFAIK Hazelcast follows this approach. They generate an error when someone is trying to launch second instance in the same working directory. Perhaps, we at least should think about changing our layout and consider including such change in some future release.
Thank you! пт, 15 мар. 2019 г. в 17:55, Dmitriy Pavlov <dpav...@apache.org>: > > Hi Ivan, yes, after the restart we need to re-use both folder name and UUID > as consistent ID. > > About returning exact directory name: I'm just guessing here, but the > persistent store is not the only folder, there are several more folders, > e.g. binary-meta, etc. > > About several nodes reusing same root folder, yes, the proposed solution > will not works because 00 is hardcoded here. But I expect it will be too > rare case. > > Sincerely, > Dmitriy Pavlov > > пт, 15 мар. 2019 г. в 11:29, Павлухин Иван <vololo...@gmail.com>: > > > Hi Dmitriy, > > > > It looks like that case with only one node in working directory is > > covered. But what is problem is solved? As I understand before fix in > > such case node after restart uses proper folder and proper > > consistentId. After fix we can specify automatically generated UUID in > > configuration and everything will go smooth. Is it the main problem > > case? If yes then such simplicity is very attractive. > > > > But honestly I do not fully understand why do we need to return a node > > to node00-UUID directory after that directory was deleted? > > > > чт, 14 мар. 2019 г. в 17:21, Dmitriy Pavlov <dpav...@apache.org>: > > > > > > Hi, I've prepared demo PR for approach 1, just checking UUID and trying > > to > > > use node00-UUID. https://github.com/apache/ignite/pull/6266/ Actually > > for > > > this approach it is really simple, no API change is required. > > > > > > I guess the case is not very often because automatic folders assignment > > > works well in most cases. > > > > > > WDYT? > > > > > > ср, 13 мар. 2019 г. в 09:42, Павлухин Иван <vololo...@gmail.com>: > > > > > > > Alex, > > > > > > > > Why do we need > > > > > - check if consistent ID is set to a string 'nodeXX-UUID'. In this > > case > > > > > the consistent ID is set to UUID, and the storage folder is chosen > > > > > according to the proper rules. This change has a minimal chance to > > affect > > > > > current users because it's unlikely that somebody is using > > auto-generated > > > > > folder naming scheme as consistent ID. > > > > ? > > > > It looks hacky as well. The thing I do not like here is that a > > > > consistentId specified in configuration is not a consistentId used by > > > > a node sometimes. > > > > > > > > Can we go just with > > > > > - either check if consistent ID is an instance of UUID and then take > > the > > > > > appropriate folder. This approach is straightforward, but may affect > > > > > current users > > > > ? > > > > > > > > And as a last chance a user will have a possibility to rename a > > directory. > > > > > > > > > > > > вт, 12 мар. 2019 г. в 19:33, Alexey Goncharuk < > > alexey.goncha...@gmail.com > > > > >: > > > > > > > > > > Igniters, > > > > > > > > > > I came across the same issue during development and found no sane > > > > > workaround for this issue. I believe the solution should be as > > simple as > > > > > possible because we are already adding a warning to let users know > > that > > > > it > > > > > is good to specify a consistent ID in production deployments. > > > > > > > > > > As for the current solution, I do not like adding a new configuration > > > > > property like 'storageFolder' because it's another way to shoot > > yourself > > > > in > > > > > the leg (e.g. different nodes have different consistent IDs but > > > > configured > > > > > to have the same storage folder). > > > > > Why can't we: > > > > > - either check if consistent ID is an instance of UUID and then > > take the > > > > > appropriate folder. This approach is straightforward, but may affect > > > > > current users > > > > > - check if consistent ID is set to a string 'nodeXX-UUID'. In this > > case > > > > > the consistent ID is set to UUID, and the storage folder is chosen > > > > > according to the proper rules. This change has a minimal chance to > > affect > > > > > current users because it's unlikely that somebody is using > > auto-generated > > > > > folder naming scheme as consistent ID. > > > > > > > > > > Thoughts? > > > > > > > > > > вт, 12 мар. 2019 г. в 16:51, Dmitriy Pavlov <dpav...@apache.org>: > > > > > > > > > > > Hi Igniters, > > > > > > > > > > > > A full description can be found at wiki page > > > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/IGNITE/Ignite+Persistent+Store+-+under+the+hood#IgnitePersistentStore-underthehood-SubfoldersGeneration > > > > > > > > > > > > I like the idea of introducing a property like nodeStorageName to > > > > specify > > > > > > exactly name of a directory to use. For example, it could have > > higher > > > > > > priority then IgniteConfiguration.getConsistentId(). It will > > impact the > > > > > > mentioned algorithm, but it seems to be easier to understand by > > users. > > > > > > > > > > > > If nobody else minds, I will try to implement this idea. > > > > > > > > > > > > Sincerely, > > > > > > Dmitriy Pavlov > > > > > > > > > > > > > > > > > > сб, 2 мар. 2019 г. в 08:33, Павлухин Иван <vololo...@gmail.com>: > > > > > > > > > > > > > Dmitriy, > > > > > > > > > > > > > > You wrote > > > > > > > > we need an order to scan and lock random-UUID based folders. > > > > > > > > > > > > > > It would be great if you provide a discussion about that order to > > > > > > > complete the picture. Currently I cannot understand why the > > order is > > > > > > > important. > > > > > > > > > > > > > > Also, couple of raw thoughts: > > > > > > > 1. Can we extend a directory lookup procedure when consistent id > > is > > > > > > > specified to check nodeXX-consistentId directories as well? > > > > > > > 2. Can we introduce a property like nodeStorageName to specify > > exact > > > > > > > name of a directory to use? It looks like a straightforward and > > > > > > > universal workaround. Or is node index better in some sense? Why? > > > > > > > > > > > > > > Please, share your thoughts. > > > > > > > > > > > > > > чт, 28 февр. 2019 г. в 17:43, Dmitriy Pavlov <dpav...@apache.org > > >: > > > > > > > > > > > > > > > > Hi Ivan, > > > > > > > > > > > > > > > > Yes, you catch me, I'm a little bit cheating with lazy > > consensus on > > > > > > code > > > > > > > > modification without providing a PR because I was expecting > > that > > > > nobody > > > > > > > > comes to discussion. I will prepare PR shortly. And since we > > anyway > > > > > > have > > > > > > > a > > > > > > > > discussion, I will not apply anything by lazy approval. > > > > > > > > > > > > > > > > - storageNodeIndex without consistent ID will not work. > > > > > > > > cfg.getDataStorageConfiguration().setNodeIdx() will be > > required > > > > only > > > > > > for > > > > > > > > case we have consistent ID. > > > > > > > > > > > > > > > > Hi Stanislav, > > > > > > > > > > > > > > > > We can't use only consistent ID because > > > > > > > > > > > > > > > > 1) we need an order to scan and lock random-UUID based folders. > > > > Node > > > > > > > index > > > > > > > > provides the order of scan. I can find the corresponding > > > > discussion, > > > > > > but > > > > > > > I > > > > > > > > guess it is not needed. > > > > > > > > 2) we need to separate backward compatible folders from new > > > > random-UUID > > > > > > > > based folders. Using UUID as folder will not allow us to scan > > only > > > > new > > > > > > > name > > > > > > > > format folders. > > > > > > > > > > > > > > > > I guess specifying node index is a quite rare case and good > > JavaDoc > > > > > > will > > > > > > > > always help. > > > > > > > > > > > > > > > > DataStorageConfiguration().setNodeIdx() JavaDoc may include > > > > following > > > > > > > > notes: > > > > > > > > Node index used for persistent store folders in case several > > nodes > > > > > > reuse > > > > > > > > one persistent store root folder. > > > > > > > > > > > > > > > > Sincerely, > > > > > > > > Dmitriy Pavlov > > > > > > > > > > > > > > > > чт, 28 февр. 2019 г. в 08:03, Павлухин Иван < > > vololo...@gmail.com>: > > > > > > > > > > > > > > > > > Dmitiy, > > > > > > > > > > > > > > > > > > Could please clarify one thing: > > > > > > > > > 1. Will it be enough to use only storageNodeIndex in order to > > > > reuse > > > > > > > > > the same persistence folders when consistentId is > > auto-generated? > > > > > > E.g. > > > > > > > > > I have a configuration with storageNodeIndex=1 and without > > > > explicitly > > > > > > > > > specified consistentId, will the node after restart use the > > same > > > > > > > > > persistence folder as before restart? > > > > > > > > > > > > > > > > > > Also a side note: > > > > > > > > > > Please share your vision. I'm going to apply this change by > > > > lazy > > > > > > > > > consensus > > > > > > > > > in 3 days. > > > > > > > > > What do you mean by "apply"? I have not seen any PR yet. > > > > > > > > > > > > > > > > > > ср, 27 февр. 2019 г. в 21:12, Dmitriy Pavlov < > > dpav...@apache.org > > > > >: > > > > > > > > > > > > > > > > > > > > Hi Igniters, > > > > > > > > > > > > > > > > > > > > I would like to fix the issue > > > > > > > > > > https://issues.apache.org/jira/browse/IGNITE-11432 about > > > > > > specifying > > > > > > > some > > > > > > > > > > previous randomly generated UUID as a new consistent ID. > > Folder > > > > > > > > > generation > > > > > > > > > > algorithm here ( > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/IGNITE/Ignite+Persistent+Store+-+under+the+hood > > > > > > > > > ) > > > > > > > > > > allows two options > > > > > > > > > > -node00+random UUID > > > > > > > > > > - consistendId > > > > > > > > > > > > > > > > > > > > I would like to add to Ignite configuration new property > > > > nodeIndex > > > > > > in > > > > > > > > > > addition to consistent Id. New Property will be named as > > > > > > > > > storageNodeIndex, > > > > > > > > > > int, zero-based. > > > > > > > > > > This will add the third option of subfolders processing: > > > > > > > > > > node{storageNodeIndex}+consistentID > > > > > > > > > > > > > > > > > > > > Please share your vision. I'm going to apply this change by > > > > lazy > > > > > > > > > consensus > > > > > > > > > > in 3 days. > > > > > > > > > > > > > > > > > > > > Sincerely, > > > > > > > > > > Dmitriy Pavlov > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > -- > > > > > > > > > Best regards, > > > > > > > > > Ivan Pavlukhin > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > -- > > > > > > > Best regards, > > > > > > > Ivan Pavlukhin > > > > > > > > > > > > > > > > > > > > > > > > > > > > > -- > > > > Best regards, > > > > Ivan Pavlukhin > > > > > > > > > > > > -- > > Best regards, > > Ivan Pavlukhin > > -- Best regards, Ivan Pavlukhin
