[WIP] Add ujson as alternative JSON encoder

[tonybaloney]

The standard library json module is the slowest of the json encoders.

ujson is 10-20x faster at encoding and decoding, especially for large datasets.

This PR moves the json imports into a shim module, which picks the standard library implementation or ujson depending on whether:

• The user has installed ujson
• The user hasn't disabled it via an environment variable

[vrdmr]

Any specific reason to choose ujon over orjson?

[tonybaloney]

Any specific reason to choose ujon over orjson?

Supporting StringifyEnum was impossible without using a fork of orjson, which I tried and it was using old bindings for Python.

ujson supports custom type serialisation via a __json__ method in the class, which is going to be more performant. It's also more compatible with json

[codecov[bot]]

Codecov Report

Merging #130 (c090330) into dev (284c15d) will decrease coverage by 0.25%. The diff coverage is 81.81%.

@@            Coverage Diff             @@
##              dev     #130      +/-   ##
==========================================
- Coverage   86.04%   85.79%   -0.26%     
==========================================
  Files          50       51       +1     
  Lines        2903     2922      +19     
  Branches      391      396       +5     
==========================================
+ Hits         2498     2507       +9     
- Misses        329      336       +7     
- Partials       76       79       +3     

Flag	Coverage Δ
unittests	85.79% <81.81%> (-0.22%)	:arrow_down:

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
azure/functions/_durable_functions.py	68.29% <ø> (ø)
azure/functions/_json.py	72.97% <72.97%> (ø)
azure/functions/_cosmosdb.py	88.88% <100.00%> (ø)
azure/functions/_http.py	91.30% <100.00%> (ø)
azure/functions/_queue.py	84.61% <100.00%> (ø)
azure/functions/_sql.py	100.00% <100.00%> (ø)
azure/functions/cosmosdb.py	74.35% <100.00%> (ø)
azure/functions/decorators/utils.py	100.00% <100.00%> (+2.53%)	:arrow_up:
azure/functions/durable_functions.py	83.33% <100.00%> (ø)
azure/functions/eventgrid.py	90.90% <100.00%> (ø)
... and 11 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update 284c15d...c090330. Read the comment docs.

[tonybaloney]

This is the benchmark between ujson (left) and json (right) for HttpRequest.get_json()

[tonybaloney]

I've deployed 2 Azure Functions in Australiaeast with this patch applied and without the patch applied

The sample POST request is:

{
	"id": "0001",
	"type": "donut",
	"name": "Cake",
	"ppu": 0.55,
	"batters":
		{
			"batter":
				[
					{ "id": "1001", "type": "Regular" },
					{ "id": "1002", "type": "Chocolate" },
					{ "id": "1003", "type": "Blueberry" },
					{ "id": "1004", "type": "Devil's Food" }
				]
		},
	"topping":
		[
			{ "id": "5001", "type": "None" },
			{ "id": "5002", "type": "Glazed" },
			{ "id": "5005", "type": "Sugar" },
			{ "id": "5007", "type": "Powdered Sugar" },
			{ "id": "5006", "type": "Chocolate with Sprinkles" },
			{ "id": "5003", "type": "Chocolate" },
			{ "id": "5004", "type": "Maple" }
		]
}

The function source code is:

import azure.functions as func
import json

def main(req: func.HttpRequest) -> func.HttpResponse:
    try:
        req_body = req.get_json()
    except ValueError:
        pass

    return func.HttpResponse(
        json.dumps(req_body),
        status_code=200
    )

The script to test the two deployments:

$ ab -p test_data.json -T application/json -n 1000 -c 10 https://ant-functions-load-testing.azurewebsites.net/api/httptriggertest
$ ab -p test_data.json -T application/json -n 1000 -c 10 https://ant-functions-load-testing-og.azurewebsites.net/api/httptriggertest

The results are:

	50	66	75	80	90	95	98	99
JSON	114	119	125	128	150	217	345	2113
UJSON	111	116	118	121	126	131	145	175
Normalised JSON	44	49	55	58	80	147	275	2043
Normalised UJSON	41	46	48	51	56	61	75	105

I've subtracted 70ms as this was the mean connect time, so you can more clearly see the difference between the two branches.

10% faster in the 50th percentile, but importantly 2.3x faster in the 95th percentile. (ignore the 99th percentile as this will include coldstart times)