[TINKERPOP-2631] GraphSON float serialization when ujson is used is imprecise

• Type: Improvement
• Status: Closed
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 3.5.1
• Fix Version/s: 3.7.0, 3.6.1, 3.5.4
• Component/s: python
• Assignee: Lyndon Bauto
• Reporter: Clément de Groc
• Created: 22/Oct/21 14:03
• Updated: 16/Jun/22 11:42
• Resolved: 16/Jun/22 11:42

Description

Problem

When submitting traversals including float values, the GraphSON-serialized traversal truncates them:

g.V().has("id", "X").property("popularity", 0.099999).iterate()

is serialized to
 

{"gremlin":{"@type":"g:Bytecode","@value":{"source":[],"step":[["V"],["has","id","X"],["property","popularity",{"@type":"g:Double","@value":0.1}],["none"]]}},"processor":"traversal"}

Cause

In gremlin_python/driver/serializer.py (but also protocol.py in passing), ujson is imported in place of stdlib json:

try:
  import ujson as json
except ImportError:
  import json

This is certainly done for performance reasons, but this has an impact on floating point precision.

With ujson < 2.0, if precise_float is not specified (which is the case):

ujson.dumps(0.099999) == '0.1'

With ujson >= 2.0, out-of-the-box floating point precision was improved: 

ujson.dumps(0.099999) == '0.099999'

Here is the change in ujson explaining the difference: ultrajson/ultrajson@eb7d894

Potential solution

Could ujson be listed as an optional dependency in setup.py with version >= 2.0 ?