Inverted-index search is constrained by the budget.

[Taewoo Kim <wa...@gmail.com>]

Hi all,

A new config parameter is added and I would like to share this with you.
The inverted-index-search in AsterixDB is different from B-Tree or R-Tree
search (they only require a certain number of frames at a time - no budget
is necessary). However, the inverted-index-search needs to read inverted
list(s) (collection of PKs for the given SK) from disk and generates
intermediate results. Previously, an inverted-index-search read the entire
inverted list into the buffer cache at once and generated intermediate
results in memory without considering any memory budget. Thus, it can
generate an OOM exception during these processes and can terminate an
instance. This problem has been fixed now. And one configuration parameter
is introduced - *compiler.textsearchmemory*. The default size is 32MB just
like other operations such as sort, group or join. If you observe an
inverted-index search is slow, try to allocate more budget to this
parameter. Thanks.

Best,
Taewoo