If the disk truly fills up HDFS should send that data to a different node, just like if you are writing to HDFS from a node that is not a compute node. However there are lots of things that can potentially go wrong if the disk truly fills up completely. Java will crash badly on startup if it's temp space if totally full. Linux behaves badly as well so much so that by default it lies about the true amount of disk space available and reserves some that only root has access to. I have not tested the rest of Map Reduce or HDFS to know for sure how well or badly they behave in this situation so there may also be other unexpected errors that happen when the disk totally fills up. Also presumably if HDFS on this one machine fills up then the other nodes are also close to full, assuming that you have run rebalancing somewhat regularly. Then The NN may have some issues trying to find free disk space and this can cause your entire cluster to slow down drastically.
--Bobby Evans On 1/10/12 4:43 AM, "Zizon Qiu" <zzd...@gmail.com> wrote: It may trigger a IOException and causing the current reduce task on that node fails,then the jobtracker will try assign that task to the other node. not quite sure. On Tue, Jan 10, 2012 at 6:19 PM, aliyeh saeedi <a1_sae...@yahoo.com> wrote: Hi I am going to save files written by reducers, but I wonder when the disk space of one node is fulfilled, what will do Hadoop? Does Hadoop put aside the node? Thank for attention
