Jacek, > We are talking here about sizes of *individual* > chunks that are transferred, my point is that > data+index are not a single file. There are two uses for this data. One is in my fault tolerance calculations from a couple of weeks back. In that case, we are assuming that one or more nodes has failed and that we need to replicate the chunks on it across all the tables. I generally expect that (a) the replicas of the failed node's chunks will be distributed across the other nodes of the cluster, although an alternative design is to have (b) another node that exactly mirrors the failed one. In case (a), the re-replication is distributed. In case (b), it's limited by the size of the output pipe on the original mirror node and the new mirror node. For fault tolerance, all data and indexes on the failed node need to be re-replicated. The second case is Mario's download-on-demand use case. (Note that I'm still not certain how realistic it will prove to be.) That use case works better with the distributed replica design. In this case, I think we may be more constrained by the outbound bandwidth of the DAC or the inbound bandwidth of the user. -- Kian-Tat Lim, LSST Data Management, [log in to unmask] ######################################################################## Use REPLY-ALL to reply to list To unsubscribe from the QSERV-L list, click the following link: https://listserv.slac.stanford.edu/cgi-bin/wa?SUBED1=QSERV-L&A=1