 So, hi Fox. So, a very good evening. So, my name is Vishal and I basically manage the Hadoop infrastructure basically at Flipkart. So, we have a lot of systems that basically depend on the Hadoop infrastructure that's basically running at Flipkart right now. So, basically I'll be speaking about the learning curves that we had during our initial HBase setup that we had initially. So, initially when we went ahead with HBase, we actually had nothing, we never had much knowledge about the internals of HBase. So, basically what happened was we went ahead with the HBase cluster setup. We went with Cloudera distribution. So, there are a couple of issues that we faced initially when we set up the cluster was, I think the foremost issue was the region servers going down because we initially went with very basic default settings which was there in HBase. So, the region servers started going down because of this and I mean it was basically hell. You had to restart region servers every now and then. So, we figured out the issue was because of major GCs happening. So, there was a term called us Juliet Paws. So, what happens is basically the region servers fail to respond and the two keeper things that they did and the two keeper just kicks them out. So, what happens is after a point in time the HBase dies off. So, this was fixed with the newer release of HBase in CDH 4.2 probably. So, what was happening was there was an MS lab configuration factor which basically dumps a GC object of a certain size. So, this actually fixed the issue of region servers going down. So, because of this what happened was we figured out every time we restarted a region server. So, it took a lot more longer because of the number of regions it had actually. So, initially when we started of when we started pumping in more data we had around 1200 regions per region server. So, what happened was each town each time a region server goes down there was this the time taken to transition each region was a lot more when compared to because of the number of regions because of the number of regions. So, to fix this issue what we did was we increased the size of each region so that the number of regions reduced considerably. So, we had said there was a configuration parameter which you guys have to set to increase the size of a particular region. So, we had set the size from 256 MB to we had increased to 50 gigs. So, what happened was we faced we started facing couple of other issues like compaction storms. So, I don't have I'm not explaining much in detail about what a compaction is in HBase. So, what happens is when a write happens it is basically fleshed into the disk and you'll have when more write happens the more flushes happen and all this files get log files get compressed or maybe compacted into a single file. So, that is what was happening when we had a large number of regions basically sorry not large number of regions of larger sizes basically. So, we had to fix this issue by looking at the region sizes. So, we later we came to we figured out that the optimal or the minimum or the optimal region size essentially should be around 10 to 15 gigs. So, that the compaction storm doesn't happen. So, even if the compaction happens it wouldn't be that frequent. So, the next parameter that we took in to focus was the by switching of compaction because compaction happens at irregular intervals. So, when you when you least needed it happens and compaction basically amplifies the read write read write that's happening in the cluster. So, we switched off compaction and I think that almost fixed the issue. So, I think we had faced a couple of a lot more issues than this. So, due to lack of time I think I would be concluding this. Thank you.