Sorting of X-Axis is incorrect when "Split Series" is also enabled

[willphillips-armedia]

Kibana version: 6.4.2

Elasticsearch version: 6.4.2

Server OS version: Windows 10 17134.345 / CentOS 6 / CentOS 7

Browser version: Internet Explorer 11.345.17134.0 / Microsoft Edge 42.17134.1.0 / Mozilla Firefox 63.0.1

Browser OS version: Windows 10 17134.345

Original install method (e.g. download page, yum, from source, etc.): Download page

Describe the bug:
Bar chart sorting of the X-Axis is incorrect when using both "X-Axis" and "Split Series" bucket options at the same time.

Steps to reproduce:

1. In Kibana, create a new Bar Chart Visualization. Select an index containing data you wish to visualize.
2. In left sidebar, select "X-Axis" bucket. Select "Terms" Aggregation and select a string field from your index. Sort and size it however you wish.
3. Click the "Apply Changes" button.
4. X-Axis bars should be sorted in the expected order.
5. Click "Add sub-buckets" button. Click "Split Series". Select "Terms" Aggregation and select a different string field from your index. Sort and size it however you wish.
6. Click the "Apply Changes" button.
7. Sometimes, data may still be ordered properly. Manipulate the amount of data through the Date Range and Query box. Typically becomes apparent when some X-Axis bars would have a zero count.

Expected behavior:

1. X-Axis ordering of bars should be consistent with or without "Split Series" bucket.
2. X-Axis ordering of bars should be consistent no matter how many records are retrieved or how they are bucketed.

Screenshots (if relevant):

Disabled "Split Series" bucket, correct ordering of X-Axis bars with custom aggregation sort

Enabled "Split Series" bucket, incorrect ordering of X-Axis bars ("31-45 Days" is shifted far-right)

Provide logs and/or server output (if relevant):

Elasticsearch Response when "Split Series" was enabled, showing ES returned data in proper order

...
{
"3": {
"doc_count_error_upper_bound": 0,
"sum_other_doc_count": 0,
"buckets": [
{
"key": "???",
"doc_count": 15
}
]
},
"key": "Within 30 Days",
"doc_count": 15,
"2-orderAgg": {
"value": 2
}
},
{
"3": {
"doc_count_error_upper_bound": 0,
"sum_other_doc_count": 0,
"buckets": [
{
"key": "???",
"doc_count": 7
}
]
},
"key": "Within 31-45 Days",
"doc_count": 7,
"2-orderAgg": {
"value": 3
}
},
{
"3": {
"doc_count_error_upper_bound": 0,
"sum_other_doc_count": 0,
"buckets": [
{
"key": "???",
"doc_count": 5
}
]
},
"key": "Within 46-60 Days",
"doc_count": 5,
"2-orderAgg": {
"value": 4
}
},
...

Any additional context:
We recently upgraded our Elastic Stack installation from a 5.X version to 6.4.2 (completely clean install, no restoration of previous indexes). Back in 5.X we noticed that "Split Series" sorting of the Legend was more of a suggestion than a consistent sort. Now, while testing 6.4.2 we have found that the X-Axis is now inconsistent as well.

What's worse, though, is that Inspecting the Elasticsearch response shows the primary aggregation buckets are returned in the correct order, thus Kibana is inconsistently disrupting the order of perfectly ordered X-Axis buckets. This is a regression from 5.X.

[elasticmachine]

Pinging @elastic/kibana-app

[timroes]

Duplicate of #17532