Hello, Thanks a lot for all these feedbacks, trying to respond to everything below:
@Parth: "I don't think we would want to maintain a TTL for the metadata store so introducing one now would mean that we might break backward compatibility down the road." Yes, I am aware of this activity starting, and I agree that whatever the solution decided later on for the new metadata store, it most probably won't support a concept of TTL. This means that we would either have to break the support of the `WITH TTL` extension of the SQL command, or to ignore it down the road. None of these solutions seem particularly appealing to me. @Padma: "What issues we need to worry about if different directories in the hierarchy are checked last at different times ?" Knowing that the refresh is always recursive, we can only have two cases: the parent level cache is refreshed at the same time as the child level cache, or the parent level cache is older than the child level cache (because a query has run in a sub-directory, that triggered a refresh of the metadata in this sub-directory). In both cases, checking the timestamp of the cache file at the root directory of the query is enough to know if the TTL criteria is respected. In the case the cache files are not refreshed at the same time between parent and children directories, and that the parent's cache is still valid with regards to its TTL, Drill would trust the parent cache, and issue an execution plan with this set of files. The same query on a child folder would use the children cache, that would have been refreshed more recently, and this would potentially result in issuing an execution plan with another set of files. So this basically, this TTL feature could create discrepancies in the results, and these discrepancies could last up to the TTL value. "Do we have to worry about coordinating against multiple drillbits ?" That would be better indeed, as the problem already exists today (I have not found any locking mechanism on metadata file), I am not sure this change would make it worse. So the reply is yes, we should worry, but I think the fix for that would be independent to this change. "Another option is if the difference between modification time of directory and metadata cache file is within TTL limit, do not do anything. If we do that, do we get errors during execution (I think so) ?" We would get errors if there would be files removed between the time of the last generation of meta-data, and the time of the execution. As in the case above, this can already happen, since there is currently no guarantee that the files at planning time will still be there at execution time. The timeframe would increase from a few milliseconds to several minutes, so the frequency of this kind of problem occurring would be much higher. I would recommend to quietly ignore missing files by considering them as empty files. "Also, how to reset that state and do metadata cache refresh eventually ?" We could reuse the REFRESH TABLE METADATA command to force the refresh. This would allow for collaborative ingestion jobs to force the refresh when the datasets have changed. Non-collaborative jobs would then rely on the TTL to get the new dataset available. "Instead of TTL, I think having a system/session option that will let us skip this check altogether would be a good thing to have. So, if we know we are not adding new data, we can set that option." I would see the need to set TTL per Table. Since different tables will have different update frequencies. I agree on a session option to bypass TTL check, so that this user will always see the last dataset. The question then becomes: what would be the default value for this option? Regards, Joel On Fri, Jul 13, 2018 at 9:06 AM, Padma Penumarthy < penumarthy.pa...@gmail.com> wrote: > Hi Joel, > > This is my understanding: > We have list of all directories (i.e. all subdirectories and their > subdirectories etc.) in the metadata > cache file of each directory. We go through that list of directories and > check > directory modification time against modification time of metadata cache > file in that directory. > If this does not match for any of the directories, we build the metadata > cache for the whole hierarchy. > The reason we have to do this adding new files will only update > modification time of immediate > parent directory and not the whole hierarchy. > > Regarding your proposal, some random thoughts: > How will you get current time that can be compared against last > modification time set by the file system ? > I think you meant compare current system time of the running java process > i.e. drillbit > against last time we checked if metadata cache needs to be updated for that > directory. > What issues we need to worry about if different directories in the > hierarchy are checked last at different times ? > Do we have to worry about coordinating against multiple drillbits ? > > Another option is if the difference between modification time of directory > and metadata cache file is within > TTL limit, do not do anything. If we do that, do we get errors during > execution (I think so) ? > Also, how to reset that state and do metadata cache refresh eventually ? > We are not saving time for modification time checks here. > > Instead of TTL, I think having a system/session option that will let us > skip this check altogether would be a > good thing to have. So, if we know we are not adding new data, we can set > that option. > > Instead of saving this TTL in metadata cache file for each > table(directory), > is it better to have this TTL as global system or session option ? > In that case, we cannot have a different TTL for each table, but it makes > it much simpler. > Otherwise, there are some complications to think about. > We have a root metadata file per directory with each of the subdirectories > underneath having their own metadata file. > So, if we update the TTL of the root directory, do we update for all the > subdirectories or just the top level directory ? > What issues we need to think about if TTL of the root directory and > subdirectories are different ? > > > Thanks > Padma > > > > > On Thu, Jul 12, 2018 at 8:07 AM, Joel Pfaff <joel.pf...@gmail.com> wrote: > > > Hello, > > > > Thanks for the feedback. > > > > The logic I had in mind was to add the TTL, as a refresh_interval field > in > > the root metadata file. > > > > At each query, the current time would be compared to the addition of the > > modification time of the root metadata file and the refresh_interval. > > If the current time is greater, it would mean the metadata may be > invalid, > > so the regular process would apply: recursively going through the file to > > check for updates, and trig a full metadata cache refresh any change is > > detected, or just touch the metadata file to align its modification time > > with the current time if no change is detected. > > If the current time is smaller, the root metadata would be trusted > (without > > additional checks) and the planning would continue. > > > > So in most of the cases, only the timestamp of the root metadata file > would > > be checked. In the worst case (at most once per TTL), all the timestamps > > would be checked. > > > > Regards, Joel > > > > On Thu, Jul 12, 2018 at 4:47 PM, Vitalii Diravka < > > vitalii.dira...@gmail.com> > > wrote: > > > > > Hi Joel, > > > > > > Sounds reasonable. > > > But if Drill checks this TTL property from metadata cache file for > every > > > query and for every file instead of file timestamp, it will not give > the > > > benefit. > > > I suppose we can add this TTL property to only root metadata cache file > > and > > > check it only once per query. > > > > > > Could you clarify the details, what is the TTL time? > > > How TTL info could be used to determine whether refresh is needed for > the > > > query? > > > > > > Kind regards > > > Vitalii > > > > > > > > > On Thu, Jul 12, 2018 at 4:40 PM Joel Pfaff <joel.pf...@gmail.com> > wrote: > > > > > > > Hello, > > > > > > > > Today, on a table for which we have created statistics (through the > > > REFRESH > > > > TABLE METADATA <path to table> command), Drill validates the > timestamp > > of > > > > every files or directory involved in the scan. > > > > > > > > If the timestamps of the files are greater than the one of the > metadata > > > > file, then a re-regeneration of the meta-data file is triggered. > > > > In the case the timestamp of the metadata file is the greatest, then > > the > > > > planning continues without regenerating the metadata. > > > > > > > > When the number of files to be queried increases, this operation can > > > take a > > > > significant amount of time. > > > > We have seen cases where this validation step alone is taking 3 to 5 > > > > seconds (just checking the timestamps), meaning the planning time was > > > > taking way more time than the querying time. > > > > And this can be problematic in some usecases where the response time > is > > > > favored compared to the `accuracy` of the data. > > > > > > > > What would you think about adding an option to the metadata > generation, > > > so > > > > that the metadata is trusted for a configurable time period > > > > Example : REFRESH TABLE METADATA <path to table> WITH TTL='15m' > > > > The exact syntax, of course, needs to be thought through. > > > > > > > > This TTL would be stored in the metadata file, and used to determine > > if a > > > > refresh is needed at each query. And this would significantly > decrease > > > the > > > > planning time when the number of files represented in the metadata > file > > > is > > > > important. > > > > > > > > Of course, this means that there could be cases where the metadata > > would > > > be > > > > wrong, so cases like the one below would need to be solved (since > they > > > may > > > > happen much more frequently): > > > > https://issues.apache.org/jira/browse/DRILL-6194 > > > > But my feeling is that since we already do have a kind of race > > condition > > > > between the view of the file system at the planning time, and the > state > > > > that will be found during the execution, we could gracefully accept > > that > > > > some files may have disappeared between the planning and the > execution. > > > > > > > > In the case the TTL would need to be changed, or be removed > completely, > > > > this could be done by re-issuing a REFRESH TABLE METADATA, either > with > > a > > > > new TTL, or without TTL at all. > > > > > > > > What do you think? > > > > > > > > Regards, Joel > > > > > > > > > >
