Fail fast on creating mutations that cannot be stored

This is in the context of Beam CloudBigtableIO. The client has a bulk mutation limit of 100k. As of #1497 this means as long as each individual Put is under that limit everything will be fine.

A Dataflow job can have large variability in the amount of data per row. It is also natural to write processElement using a single Put, especially since that is what the examples guide new users to do.

If a single Put is larger than 100k mutations then it is impossible to store. The exception it causes is thrown later, and from a different thread.

The client should throw the TooManyMutations exception much earlier, preferably so the user code that caused the 100001th mutation it is still in the stack trace.

[sduskis]

We definitely show throw exceptions earlier. We can also add an option (with a big Caveat Emptor) to allow multiple Puts in this case.

[sduskis]

@alugowski, thanks for the context.

FYI, Put is owned by HBase, which doesn’t have the 100K limitation, so we couldn’t add an exception there. We added the warning in our implementation of BufferedMutator.mutate, which is called under the cover in CloudBigtableIO.

We were exploring possibilities of how to improve the experience. A note in BigQueryBigtableTransfer would help. We may also fix our import job, where users can’t control the code themselves.