Reindex: Remove ability to sort

[henningandersen]

As part of the reindex job specification sorting can be specified. Documentation describes that this can be used in combination with max_docs to extract either a specific or a random subset of data.

However, specifying sorting is not compatible with the new upcoming resilient reindex mechanism, since this relies on sorting by seq_no. Any reindex request that sorts by anything but seq_no first will not be resilient.

When copying the full data set, sorting does not really make a difference, the net end result will be the same. Extracting subsets of data can likely be done by adding queries instead. To avoid having cases where reindex is not resilient, I propose to deprecate sorting in reindex in 7.x and remove sorting from reindex in 8.0.

This issue is created to gather feedback on this proposal. If you rely on being able to sort while reindexing, please let us know here.

[elasticmachine]

Pinging @elastic/es-distributed (:Distributed/Reindex)

[henningandersen]

We discussed this in our FixitFriday meeting and concluded that we think removing the ability to sort in reindex is the right path forward. Following is a summary of the use cases discussed that might be affected and their possible workarounds:

1. Resilience
   1. Sort might be useful to add resilience to reindex.
   2. Most such cases would directly do search/scroll and bulk requests since it gives more control.
   3. Resilience will become native to reindex in the future. We should ensure that resilient reindex is a reality before completely removing the sort option.
2. Incremental reindex
   1. Sorting by a timestamp (or _seq_no) field might be used for providing incremental reindex, trying to only index newly arrived docs since last reindex.
   2. An alternative is to do range based reindex requests instead.
   3. A separate alternative is to handle the search/scroll and bulk outside ES instead.
   4. It should be noted that handling deletes on source will likely require an external program doing bulk requests anyway.
3. Get last X docs.
   1. This is a documentation example and thus not necessarily a valid use case.
   2. Getting a specific subset of docs could be expressed as a range filter instead.
4. Get a random subset of docs
   1. This is a documentation example and thus not necessarily a valid use case.
   2. A similar result can be obtained without sorting by adding a suitable min_score to the function_score clause, for instance "min_score" : 0.9.

The plan is to move forward with deprecating the ability to sort during reindex. We will give this another 2-3 weeks here to gather input before taking concrete actions. Our current plan after that is to:

1. Deprecate sorting in reindex in 7.x such that deprecation warnings are emitted if used. We prefer to deprecate early to maximize the period where users could run into the deprecation warning.
2. Remove the documentation examples using sort in reindex from master.
3. When resilient reindex has reached feature parity with current reindex, remove the ability to sort in master (breaking change).