PR: https://github.com/OpenNMS/nephron/pull/31

Pane accumulation alone was not enough. In case of early / late results it can happen that an early result pane of a later window completes before a late result pane of an earlier window.

A stateful transform in the global window is used to assign unique index numbers to results for the same event timestamp. That index is used as an additional metrics label in Cortex.

Cortex requires that metric values are written ordered by timestamps (i.e. older values must not be written after newer values have been written). Beam however, does not guarantee processing order, i.e. windows (and panes) may be processed in parallel and "older" results may complete earlier than "newer" results.

I implemented a stateful PaneAccumulator transformation that can accumulate values in several panes of the same window into one or more output values. Accumulated values are tupled with a sequential index number that can be used as an additional label for cortex metrics. When metrics are retrieved cortex can sum over all metrics that differ in that index only.

It has to be tested how PaneAccumulator influences performance. There is some additional state to be stored per window/key and some additonal shuffling done.