ctx._now time seems to be incorrect

[seang-es]

To repro:

Create a document, then update it as follows:

POST my-index/my-type/1/_update
{
   "script" : {
       "inline": "ctx._source.timestamp_now = ctx._now; ctx._source.epochms = new Date().getTime();",
       "lang": "painless"
   },
   "upsert" : {
       "foo":"bar"
   }
}

In the results we're seeing, epochms is correct, but timestamp_now seems to be a random time. The ctx._now time is consistently increasing, but on various different nodes it has returned times ranging from 2003 to 2019.

[rjernst]

@seang-es Can you please include the exact mappings, as well as the output when retrieving the updated document?

[jpcarey]

Elasticsearch 5.2.0 + x-pack, macOS Sierra 10.12.3 (16D32), Oracle java version "1.8.0_65"

POST t/t/1
{
  "foo": "bar"
}
POST t/t/1/_update
{
  "script": {
    "inline": "ctx._source.timestamp_now = ctx._now; ctx._source.epochms = new Date().getTime();",
    "lang": "painless"
  },
  "upsert": {
    "foo": "bar2"
  }
}

GET t/t/1 Response:

{
  "_index": "t",
  "_type": "t",
  "_id": "1",
  "_version": 4,
  "found": true,
  "_source": {
    "foo": "bar",
    "timestamp_now": 1206909990,
    "epochms": 1487111464024
  }
}

GET t/_mapping Response:

{
  "t": {
    "mappings": {
      "t": {
        "properties": {
          "epochms": {
            "type": "long"
          },
          "foo": {
            "type": "text",
            "fields": {
              "keyword": {
                "type": "keyword",
                "ignore_above": 256
              }
            }
          },
          "timestamp_now": {
            "type": "long"
          }
        }
      }
    }
  }
}

[s1monw]

I think this is a bug. _now is set to nanotime in millis.

[rjernst]

I found the bug. We currently use System.nanoTime() and convert that to milliseconds, but that is not milliseconds since epoch. This was changed about 1.5 years ago to avoid System.currentTimeMillis() from going backwards (see #11626). We either need to go back to System.currentTimeMillis(), or look at how to protecting from negative time diffs differently.

[rjernst]

Oops sorry Simon, I guess that's what you were trying to say. :)