39th European Conference on Information Retrieval Research (ECIR), Aberdeen, United Kingdom, 8 - 13 April 2017, vol.10193, pp.705-712
Our work shows that the query latency for selective search over a topically partitioned collection can be reduced by up to 55%. We achieve this by physically storing the documents in each topical cluster across all shards and building a cluster-skipping index at each shard. Our approach also achieves uniform load balance among the shards.