Saw an interesting hack when a colleague sent me his code for review. We are a cloud storage company and people upload files and if they upload same file again and again it creates a new version. Some examples are quickbooks or outlook files that will generate multiple versions in a day if you have enabled real time sync on that folder where these files are stored. To optimize bandwidth we use rsync and do a patch on server to reconstruct these large files but as we save the original file the customer gets charged the full size of the file. This is why customers configure version policy that they would allow 5 versions of the file and if a new one is uploaded we move the oldest to trash. Now if the customer reduced the versions to keep from 5 to 2 then suddenly we have to delete all these versions.
So earlier to offload processing we had written a rest api that in streaming fashion would return list of deletable versions metadata.
/rest/public/getDeletableVersions GET.
then a python script would call deleteVersions api in batch.
/rest/public/deleteVersions POST
now this was all complex to test so after 2 years we moved it back to tomcat and rewrote this as a quartz job.
so the programmer reused the code and wrote it as
List<DeletableVersionResponse> deletableVersions = storageService.getDeletableVersions(customerId);
for(List<DeletableVersionResponse> batch: split(deletableVersions) ) {
storageService.deleteVersions(batch) ;
}
Problem is that for bigger customers that had 10M+ versions this was causing OOM when we were trying to load all deletable versions.
So I asked the engineer to convert it in such a manner that the api would be
int numDeleted;
while((numDeleted=storageService.deleteNextBatchOfDeletableVersions(customerId))>0) {
}
but this required change in all the layers.
Instead the engineer came up with solution to use BlockingQueue.
So what he did was
BlockingQueue<DeletableVersionResponse> deletableVersionResponse = new ArrayBlockingQueue<DeletableVersionResponse>(
eventBatchSize);
Future future = executorService.submit(new DeletableVersionRequest(deletableVersionResponse, customerId));
while (!(deletableVersionResponse.isEmpty() && future.isDone())) {
processDeletableVersionResponse(user, deletableVersionResponse.poll());
}
This was a creative way to solve OOM without changing a lot of layers of code.
So earlier to offload processing we had written a rest api that in streaming fashion would return list of deletable versions metadata.
/rest/public/getDeletableVersions GET.
then a python script would call deleteVersions api in batch.
/rest/public/deleteVersions POST
now this was all complex to test so after 2 years we moved it back to tomcat and rewrote this as a quartz job.
so the programmer reused the code and wrote it as
List<DeletableVersionResponse> deletableVersions = storageService.getDeletableVersions(customerId);
for(List<DeletableVersionResponse> batch: split(deletableVersions) ) {
storageService.deleteVersions(batch) ;
}
Problem is that for bigger customers that had 10M+ versions this was causing OOM when we were trying to load all deletable versions.
So I asked the engineer to convert it in such a manner that the api would be
int numDeleted;
while((numDeleted=storageService.deleteNextBatchOfDeletableVersions(customerId))>0) {
}
but this required change in all the layers.
Instead the engineer came up with solution to use BlockingQueue.
So what he did was
BlockingQueue<DeletableVersionResponse> deletableVersionResponse = new ArrayBlockingQueue<DeletableVersionResponse>(
eventBatchSize);
Future future = executorService.submit(new DeletableVersionRequest(deletableVersionResponse, customerId));
while (!(deletableVersionResponse.isEmpty() && future.isDone())) {
processDeletableVersionResponse(user, deletableVersionResponse.poll());
}
This was a creative way to solve OOM without changing a lot of layers of code.
Comments
Post a Comment