Large Survey Database: A Distributed Framework for Storage and Analysis of Large Datasets

The Large Survey Database (LSD) is a Python framework and DBMS for distributed storage, cross-matching and querying of large survey catalogs (>10^9 rows, >1 TB). The primary driver behind its development is the analysis of Pan-STARRS PS1 data. It is specifically optimized for fast queries and parallel sweeps of positionally and temporally indexed datasets. It transparently scales to more than >10^2 nodes, and can be made to function in "shared nothing" architectures.

An LSD database consists of a set of vertically and horizontally partitioned tables, physically stored as compressed HDF5 files. Vertically, we partition the tables into groups of related columns ('column groups'), storing together logically related data (e.g., astrometry, photometry). Horizontally, the tables are partitioned into partially overlapping ``cells'' by position in space (lon, lat) and time (t). This organization allows for fast lookups based on spatial and temporal coordinates, as well as data and task distribution. The design was inspired by the success of Google BigTable (Chang et al., 2006).

Our programming model is a pipelined extension of MapReduce (Dean and Ghemawat, 2004). An SQL-like query language is used to access data. For complex tasks, map-reduce ``kernels'' that operate on query results on a per-cell basis can be written, with the framework taking care of scheduling and execution. The combination leverages users' familiarity with SQL, while offering a fully distributed computing environment.

LSD adds little overhead compared to direct Python file I/O. In tests, we sweeped through 1.1 Grows of PanSTARRS+SDSS data (220GB) less than 15 minutes on a dual CPU machine. In a cluster environment, we achieved bandwidths of 17Gbits/sec (I/O limited). Based on current experience, we believe LSD should scale to be useful for analysis and storage of LSST-scale datasets. It can be downloaded from http://mwscience.net/lsd.