Consider the case of computing average temperature, grouped by city.

For this example, let's assume that this is the original data set and that it is initially partitioned into two subsets as follows:

In the first step, we compute partial sum and count for each of the data partitions as follows:

Note that in this example, the partitions output one row for each group. Depending on how many rows there are and how much memory is available, each partition may output multiple rows per group. The Group operator maintains an LRU cache of partials, keyed by group key so as to make a "best effort" to reduce the amount of data to redistribute in the next step.

Next, we redistribute by city such that all partials of the same city live within the same partition. The repartitioning is somewhat arbitrary as long as a city does not span partitions. For this example, we'll assign Boston and Austin to partition 1 and Seattle and San Francisco to partition 2. Furthermore, within each partition, we sort by city.

Note that the framework will skip this step if the input data is already distributed/sorted by the group keys.

Next, combine partials by group:

Finally, we compute average by dividing sums by counts: