JsonSerializer reading from Utf8JsonReader

[BrennanConroy]

Currently if you want to get an object out of a Utf8JsonReader you need to first use the reader with a temporary JsonDocument, then pass the JsonElement.GetRawText() from that document into the JsonSerializer.

There should be a way to do this directly from the serializer and avoid the temporary JsonDocument and GetRawText() call.

var reader = new Utf8JsonReader(someBytes, isFinalBlock: true, state: default);
using var token = JsonDocument.ParseValue(ref reader);
var obj = JsonSerializer.Parse(token.RootElement.GetRawText(), someType);

The requested API would look something like:

var reader = new Utf8JsonReader(someBytes, isFinalBlock: true, state: default);
var obj = JsonSerializer.Parse(ref reader, someType);

I did an perf test with the current approach vs. something similar to the proposed change

TLDR

Using something similar to JsonSerializer.Parse(ref reader, someType); could improve Op/s from 22,587 to 59,662 and allocations from 50.28KB to 20.23KB in the scenario I was testing.

Current Approach:

Method	Input	HubProtocol	Mean	Error	StdDev	Op/s	Gen 0	Allocated
ReadSingleMessage	LargeArguments	Json	44.27 us	1.3590 us	4.007 us	22,587.6	1.4648	50.28 KB
ReadSingleMessage	LargeArguments	NewtonsoftJson	51.58 us	1.0247 us	2.552 us	19,386.1	0.5493	20.68 KB

Using the original someBytes and reader.BytesConsumed to strategically slice data off and pass it directly to JsonSerializer.Parse(...):

Method	Input	HubProtocol	Mean	Error	StdDev	Op/s	Gen 0	Allocated
ReadSingleMessage	LargeArguments	Json	16.76 us	0.4949 us	1.459 us	59,662.9	0.5493	20.23 KB
ReadSingleMessage	LargeArguments	NewtonsoftJson	51.86 us	1.0326 us	1.725 us	19,282.0	0.5493	20.68 KB

[ahsonkhan]

NewtonsoftJson 51.58 -> 39.31. Why did that one change?

[BrennanConroy]

I'm not sure, that one confused me too. I ran it twice and got the same result.

Maybe benchmark dotnet isn't as sandboxed as it could be

---

From: Ahson Khan notifications@github.com Sent: Monday, April 8, 2019 6:01:24 PM To: dotnet/corefx Cc: Brennan Conroy; Author Subject: Re: [dotnet/corefx] JsonSerializer reading from Utf8JsonReader (#36717)

NewtonsoftJson 51.58 -> 39.31. Why did that one change?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHubhttps://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fdotnet%2Fcorefx%2Fissues%2F36717%23issuecomment-481061469&data=02%7C01%7Cbrecon%40microsoft.com%7Ca680792d94134157a92508d6bc86e34e%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636903684858807240&sdata=WgvpAyb7HD%2BYKIbF29%2BZKEa%2B63upsIGS6m%2BSJxechSo%3D&reserved=0, or mute the threadhttps://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAHOVEXFKSOC5081vHYd1QR_4THSqszwWks5ve-ZkgaJpZM4cjWJe&data=02%7C01%7Cbrecon%40microsoft.com%7Ca680792d94134157a92508d6bc86e34e%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636903684858817248&sdata=0EHeQf9%2BRqjPAanhFG7SgMd5vcCo0bVzPnTk9s7VDjM%3D&reserved=0.

[BrennanConroy]

Tried it again and got more reasonable numbers. Updated inline

[stephan-tolksdorf]

Such a Parse overload would allow to e.g. deserialize a params object in a JSON RPC request without having to read (and validate) the JSON representation of the params object twice.

[rynowak]

We have a use case for this when writing custom converters. Frequently you want to use the JsonSerializer to read a value from the reader so that the serializer can do the right thing.

public static partial class JsonSerializer
{
    public static object ReadValue(ref Utf8JsonReader reader, Type returnType, JsonSerializerOptions options = null) { throw null; }
    public static TValue ReadValue<TValue>(ref Utf8JsonReader reader, JsonSerializerOptions options = null) { throw null; }
}

[steveharter]

The ReadValue terminology -- does just Read work? We have ReadAsync already and that would be consistent.

Design notes:

• The buffer in the reader must contain all of the data necessary to return the value - i.e. the reader shouldn't return false to request more data.
• If there is extra data in the buffer after deserialization that is OK - the deserializer should not throw. This is different than the normal Parse methods which don't allow additional data. This also requires the serializer to track the state, or inspect the reader's internal state to determine the current json depth.

[ahsonkhan]

The ReadValue terminology -- does just Read work? We have ReadAsync already and that would be consistent.

The ReadAsync tries to read as much as possible. ReadValue would be consistent with JsonDocument.ParseValue which reads just one value (whether it be an object/array or primitive).

[terrajobst]

Video

Approved because we approved the combined one dotnet/runtime#29205.

[BrennanConroy]

Did a before and after for this change: Before:

Method	Input	HubProtocol	Mean	Error	StdDev	Op/s	Gen 0	Gen 1	Allocated
ReadSingleMessage	LargeArguments	Json	21.11 us	1.6846 us	4.967 us	47,381.8	2.0752	0.0610	50.29 KB
ReadSingleMessage	LargeArguments	NewtonsoftJson	19.43 us	0.4236 us	1.236 us	51,469.2	0.7019	-	20.68 KB

After:

Method	Input	HubProtocol	Mean	Error	StdDev	Op/s	Gen 0	Allocated
ReadSingleMessage	LargeArguments	Json	9.199 us	0.5671 us	1.672 us	108,709.4	0.6714	20.15 KB
ReadSingleMessage	LargeArguments	NewtonsoftJson	19.627 us	0.3924 us	1.126 us	50,950.9	0.7019	20.68 KB

With custom slicing code (near optimal perf):

Method	Input	HubProtocol	Mean	Error	StdDev	Op/s	Gen 0	Allocated
ReadSingleMessage	LargeArguments	Json	8.244 us	0.5701 us	1.672 us	121,294.4	0.6866	20.23 KB
ReadSingleMessage	LargeArguments	NewtonsoftJson	20.251 us	0.3925 us	1.088 us	49,381.2	0.7019	20.68 KB