118 Big Data Computing
becomes highly nontrivial. Furthermore, data provenance is a total night-
mare. It’s just a really ugly architecture to have these two systems with a
connector between them.”
Michael J. Carey adds that is:
• Questionable to layer a record-oriented data abstraction on top of a
giant globally sequenced byte-stream le abstraction.
(E.g., HDFS is unaware of record boundaries. “Broken records” instead of
xed-length le splits, i.e., a record with some of its bytes in one split and
some in the next)
• Questionable building a parallel data runtime on top of a unary
operator model (map, reduce, combine). E.g., performing joins with
MapReduce.
• Questionable building a key-value store layer with a remote query
access ...