Pinging @elastic/ml-core (:ml)

Given what the C++ does - see https://github.com/elastic/ml-cpp/pull/584/files#diff-8c1b056d827cb41f2dbd4a97a5e202deR59 - it might be better to change DataFrameAnalyticsConfig.MIN_MODEL_MEMORY_LIMIT to 1kb instead of messing with the estimates in the "estimate" action. If we cap at 1mb here then there is no point in the C++ outputting values that are less than a kilobyte.

I didn't want to change C++ as it's not the layer that has the knowledge of this minimum memory limit.

I am pretty sure we once discussed whether model memory limits for data frame analytics jobs should be whole numbers of megabytes like they are for anomaly detection jobs. I cannot find the whole discussion now. It was touched on in elastic/ml-cpp#584 (comment), but I think there was a bigger discussion somewhere where there was a desire to allow model memory limits less than a megabyte for really small jobs.

I'm all for lowering the limit to 1kB. Having said that, I don't remember the outcome of the discussion you are referring to so I may be missing some context.

Changing the min model memory limit to 1kb might have side other effects though, so please check how it's used throughout the code.

The minimum limit of 1MB is only used during config creation to reject configs for which the limit is set lower.
Looking at where the DataFrameAnalyticsConfig.getModelMemoryLimit() is used, I cannot see a direct threat to jobs functioning.

Code LGTM, but please can you change the PR title and description before merging

Done.