Deepsort/deterministic toJson Mustache function

[ypid-geberit]

Describe the feature:

The Mustache function {{#toJosn}} is not deterministic meaning that the order of keys in a dictionary can change when templating the same data multiple times. This is an issue for unit/integration testing where exact matches are required. On ES 6.2.4 I never had those issues. Now on ES 6.5.1 dictionaries are serialized not deterministical.

I am surprised that this has not caused issues for the unit tests of Elasticsearch itself. Example: https://github.com/elastic/elasticsearch/pull/18856/files#diff-ba0eba6caacd2ed288fbac13b74086eaR239

Maybe it only happens for more nested data structures. I found the issue while writing an Elasticsearch watch like this:

---

# yamllint disable rule:line-length rule:comments-indentation

metadata:

  ## Use an offset to allow events to travel thought their pipeline.
  ## See: https://discuss.elastic.co/t/ensure-that-watcher-does-not-miss-documents-logs/127780/1
  time_offset: '5m'
  time_window: '1m'

trigger:
  schedule:
    ## Based on ctx.metadata.time_window.
    ## Everyone triggers jobs at second 0 so this one is triggered at second 50
    ## every minute.
    cron:
      - '50 * *    * * ?'

input:
  search:
    ## Based on ctx.metadata.time_window.
    timeout: '40s'
    request:

      indices:
        - 'log_network-switch__*'

      body:

        size: 100

        query:
          bool:
            filter:
              - match_phrase:
                  host:
                    query: 'gnu'

condition:
  compare:
    ctx.payload.hits.total:
      gt: 0

actions:
  log:
    logging:
      level: info
      text: 'Watch query hits: {{#toJson}}ctx.payload.hits.hits{{/toJson}}'

(Yes, YAML is awesome, also for watch definitions. Ref: elastic/examples#239)

Maybe also other parameters like pretty, indent could be passed to {{#toJosn}} similar to {{#join}}. Ref: #18856

I at first used:

      text: 'Watch query hits: {{ctx.payload.hits.hits}}'

but this is also not deterministic.

[elasticmachine]

Pinging @elastic/es-core-infra

[danielmitterdorfer]

We discussed this in our weekly team meeting. While it would be technically possible to do this, we don't want to encourage to rely on ordering in JSON content. Therefore, I am closing it.

[ypid-geberit]

How else should users of the integration testing of Elasticsearch watches (https://github.com/elastic/examples/tree/master/Alerting/Sample%20Watches) test complex watches then? With this decision, I am forced to write my own deterministic data serialization function in Painless.

Can you give more reasoning about this decision? My idea is that the underlying JSON library you are using supports this already and you only need to pass a parameter like "pretty=true" or something. Without this, it would require some custom JSON implementation for serialization. Currently my workaround is to delete all nested hashes for testing, this is enough for me currently but I fear the situation where I also need to test deep hash datasets.