Remove String conversion in decoders. #671

rpless · 2016-11-19T02:35:23Z

Proposed fix for #511. The only thing I was unsure about was whether to use Buf.ByteArray.Owned or Buf.ByteArray.Shared for the libraries that can convert Array[Byte]. I wound up using Shared, but I'm not sure this is necessary.

codecov-io · 2016-11-19T02:41:14Z

Current coverage is 74.34% (diff: 76.00%)

Merging #671 into master will decrease coverage by 0.32%

@@             master       #671   diff @@
==========================================
  Files            30         30          
  Lines           592        608    +16   
  Methods         566        586    +20   
  Messages          0          0          
  Branches         26         22     -4   
==========================================
+ Hits            442        452    +10   
- Misses          150        156     +6   
  Partials          0          0

Powered by Codecov. Last update 2d479db...a5335d1

vkostyukov · 2016-11-19T02:58:43Z

@rpless Thanks a lot for looking into this!

Some suggestions:

Instead of Buf.ByteArray.Shared (which always copy), let's use this trick that allows retrieving a byte array from Netty's channel buffer. This heavily depends on the underlying implementation, but I'm fine with that for now.
Let run some benchmarks (both JMH and wrk) to see how this affects throughput (I'm excited to see the numbers!)

vkostyukov · 2016-11-19T06:15:46Z

Trying to be more specific here.

If you need ByteBuffer, the cheapest way (at least for Netty 3) to get it is:

val byteBuffer = ChannelBufferBuf.Owned.extract(buf).toByteBuffer()

If you need Array[Byte], the cheapest way to get it is:

val channelBuffer = ChannelBufferBuf.Owned.extract(buf)
val (byteArray, offset, length) = 
  // assert channelBuffer.hasArray()
  (channelBuffer.array(), channelBuffer.readerIndex(), channelBuffer.readableBytes())

rpless · 2016-11-19T13:48:09Z

I ran the wrk benchmarks for Finch + circe. The results are here.
As for jmh benchmarks, unless I missed something, it doesn't look like we have ones that exercise the the json decoders anymore.

vkostyukov · 2016-11-19T16:56:33Z

Nice!

@rpless Do you mind keep the server running and run wrk for 3 times. Usually, it takes some time for JVM to warmup so using the results from 3rd run seems like the closest we can get to the steady state.

vkostyukov

Just a few nits from me. Really nice job!

vkostyukov · 2016-11-19T16:57:37Z

argonaut/src/main/scala/io/finch/argonaut/Decoders.scala

@@ -1,6 +1,8 @@
 package io.finch.argonaut

 import argonaut.{CursorHistory, DecodeJson, Json}
+import com.twitter.finagle.netty3.ChannelBufferBuf
+import com.twitter.io.Charsets


Please, use StandardCharsets from JDK. c.t.i.Charsets are deprecated in the most recent Finagle release.

vkostyukov · 2016-11-19T16:58:30Z

circe/src/main/scala/io/finch/circe/Decoders.scala

-  )
+  implicit def decodeCirce[A: Decoder]: Decode.Json[A] = Decode.json({ (b, cs) =>
+    val attemptJson = cs match {
+      case Charsets.Utf8 => parseByteBuffer(ChannelBufferBuf.Owned.extract(b).toByteBuffer()).right.flatMap(_.as[A])


Is scalacheck happy with that line length?

Yes, its under the line limit by 3 character. Happy to move it down a line if you think that's more readable.

Nah, it's fine. Thanks!

rpless · 2016-11-19T17:50:06Z

Updated benckmarks, results after 3 runs:

Without String Conversion (this PR)

$ wrk -t4 -c24 -d30s -s examples/src/main/scala/io/finch/wrk/wrk.lua http://localhost:8081/
Running 30s test @ http://localhost:8081/
  4 threads and 24 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     1.68ms    3.67ms  52.43ms   90.18%
    Req/Sec    12.20k     2.26k   23.58k    67.25%
  1458406 requests in 30.09s, 137.69MB read
Requests/sec:  48471.67
Transfer/sec:      4.58MB

With String Conversion (Current Finch)

$ wrk -t4 -c24 -d30s -s examples/src/main/scala/io/finch/wrk/wrk.lua http://localhost:8081/
Running 30s test @ http://localhost:8081/
  4 threads and 24 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     1.62ms    3.22ms  62.64ms   89.91%
    Req/Sec    10.84k     2.44k   22.20k    68.33%
  1296147 requests in 30.10s, 122.37MB read
Requests/sec:  43067.06
Transfer/sec:      4.07MB

vkostyukov · 2016-11-19T17:54:09Z

Looks like 12.5% improvement in the throughput! Sweet!

clhodapp · 2016-11-19T18:56:52Z

playjson/src/main/scala/io/finch/playjson/package.scala

+      val buf = ChannelBufferBuf.Owned.extract(b)
+      if (buf.hasArray) Try(Json.parse(new String(buf.array(), 0, buf.readableBytes(), cs)).as[A])
+      else Try(Json.parse(buf.toString(cs)).as[A])
+    })


It looks like you might be parsing twice here by accident?

No accident. Its based on this trick that @vkostyukov mentioned in earlier. There may be a better to express this (i.e. only having on call to Json.parse once and having the if/ else only return a Try[String]).

@clhodapp I played around with it a little. Do you think this version makes the intent more clear?

val buf = ChannelBufferBuf.Owned.extract(b) val bufAsString = Try( if (buf.hasArray) new String(buf.array(), 0, buf.readableBytes(), cs) else buf.toString(cs) ) bufAsString.map(Json.parse(_).as[A])

vkostyukov · 2016-11-19T19:16:44Z

I think Chris is right. We call Json.parse two times and drop the first
result on the floor.

On Sat, Nov 19, 2016 at 11:14 AM Ryan Plessner notifications@github.com
wrote:

@rpless commented on this pull request.

In playjson/src/main/scala/io/finch/playjson/package.scala
#671:
@@ -11,10 +13,11 @@ package object playjson {
* @tparam A the type of the data to decode into
*/
implicit def decodePlayJson[A](implicit reads: Reads[A]): Decode.Json[A] =
// TODO: Eliminate toString conversion

// See Eliminate meaningless to-string conversion #511

// PlayJson can parse from byte[] automatically detecting the charset.

Decode.json((b, cs) => Try(Json.parse(BufText.extract(b, cs)).as[A]))

Decode.json({ (b, cs) => Try(Json.parse(Buf.ByteArray.Shared.extract(b)).as[A])
 val buf = ChannelBufferBuf.Owned.extract(b)
 if (buf.hasArray) Try(Json.parse(new String(buf.array(), 0, buf.readableBytes(), cs)).as[A])
 else Try(Json.parse(buf.toString(cs)).as[A])
})
@clhodapp https://github.com/clhodapp I played around with it a little.
Do you think this version makes the intent more clear?

val buf = ChannelBufferBuf.Owned.extract(b)val bufAsString = Try(
if (buf.hasArray) new String(buf.array(), 0, buf.readableBytes(), cs)
else buf.toString(cs)
)
bufAsString.map(Json.parse(_).as[A])

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#671, or mute the thread
https://github.com/notifications/unsubscribe-auth/ABDDKwEmt9Qm3tSF-W2xoHHcsjNTIvoyks5q_0p5gaJpZM4K3Isa
.

rpless · 2016-11-19T19:19:44Z

Ah I see it now. The line above is the old code. I'll remove it.

vkostyukov · 2016-11-19T19:29:23Z

playjson/src/main/scala/io/finch/playjson/package.scala

-    Decode.json((b, cs) => Try(Json.parse(BufText.extract(b, cs)).as[A]))
+    Decode.json({ (b, cs) =>
+      val buf = ChannelBufferBuf.Owned.extract(b)
+      if (buf.hasArray) Try(Json.parse(new String(buf.array(), 0, buf.readableBytes(), cs)).as[A])


I think, Json.parse can parse byte[] directly. No need for new String.

Json.Parse can take a byte[], but when I convert that line to Try(Json.parse(buf.array()).as[A]) it gives me the following in the

Throw(spray.json.JsonParser$ParsingException: Unexpected character '\u0000'

I think I'm missing something about how the ChannelBufferBuf holds onto the underlying array.

vkostyukov · 2016-11-19T19:29:39Z

sprayjson/src/main/scala/io/finch/sprayjson/package.scala

+      cs match {
+        case StandardCharsets.UTF_8 =>
+          val buf = ChannelBufferBuf.Owned.extract(b)
+          if (buf.hasArray) Try(JsonParser(new String(buf.array(), 0, buf.readableBytes(), cs)).convertTo[A])


No need for new String.

A similar issue exists here. The tests raise this error.

Throw(com.fasterxml.jackson.core.JsonParseException: Illegal character ((CTRL-CHAR, code 0)): only regular white space (\r, \n, \t) is allowed between tokens

vkostyukov · 2016-11-19T19:30:21Z

circe/src/main/scala/io/finch/circe/Decoders.scala

-      value => Return(value)
-    )
-  )
+  implicit def decodeCirce[A: Decoder]: Decode.Json[A] = Decode.json({ (b, cs) =>


Let's remove wrapping (). Just Decode.json { ... } is fine.

vkostyukov · 2016-11-19T19:30:48Z

jackson/src/main/scala/io/finch/jackson/package.scala

-    // Jackson can parse from byte[] automatically detecting the encoding.
-    Try(mapper.readValue(BufText.extract(b, cs), ct.runtimeClass.asInstanceOf[Class[A]]))
-  )
+  ): Decode.Json[A] = Decode.json({ (b, cs) =>


Please, remove (): Decode.json { ... }.

vkostyukov · 2016-11-19T19:31:08Z

playjson/src/main/scala/io/finch/playjson/package.scala

-    // See https://github.com/finagle/finch/issues/511
-    // PlayJson can parse from byte[] automatically detecting the charset.
-    Decode.json((b, cs) => Try(Json.parse(BufText.extract(b, cs)).as[A]))
+    Decode.json({ (b, cs) =>


Please, remove wrapping (): Decode.json { ... }.

vkostyukov · 2016-11-19T19:31:31Z

sprayjson/src/main/scala/io/finch/sprayjson/package.scala

-    // See https://github.com/finagle/finch/issues/511
-    // SprayJson can parse from byte[] if it represents a UTF-8 string.
-    Decode.json((b, cs) => Try(BufText.extract(b, cs).parseJson.convertTo[A]))
+    Decode.json({ (b, cs) =>


Please, remove wrapping (): Decode.json { ... }.

rpless · 2016-11-19T20:48:42Z

Ok, it looks like using slice like this works: buf.array().slice(0, buf.readableBytes()). Does that make sense?

vkostyukov · 2016-11-19T20:59:03Z

Slice works but it allocates a new array. You should be able to bypass
readableBytes to the parse function.

On Sat, Nov 19, 2016 at 12:48 PM Ryan Plessner notifications@github.com
wrote:

Ok, it looks like using slice like this works: buf.array().slice(0,
buf.readableBytes()). Does that make sense?

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#671 (comment), or mute
the thread
https://github.com/notifications/unsubscribe-auth/ABDDK-0w2n0qAoIwWMcWBZ8h16aUD1ooks5q_2CrgaJpZM4K3Isa
.

rpless · 2016-11-19T21:09:45Z

Yeah, I'm not a fan of the new array allocation, but it seems like neither Spray Json nor Play Json expose a parsing function that takes the number of bytes to read:

vkostyukov · 2016-11-19T21:17:00Z

Play takes InputStream though. Can we wrap our byte array with
ByteArrayInputStream and specify the length there?

For Spray, let's do System.arraycopy it's still better than 'new String'.

On Sat, Nov 19, 2016 at 1:10 PM Ryan Plessner notifications@github.com
wrote:

Yeah, I'm not a fan of the new array allocation, but it seems like neither
Spray Json nor Play Json expose a parsing function that takes the number of
bytes to read:

https://www.playframework.com/documentation/2.5.x/api/scala/index.html#play.api.libs.json.Json$

https://github.com/spray/spray-json/blob/master/src/main/scala/spray/json/JsonParser.scala#L28
(and also
https://github.com/spray/spray-json/blob/master/src/main/scala/spray/json/JsonParser.scala#L237
)

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#671 (comment), or mute
the thread
https://github.com/notifications/unsubscribe-auth/ABDDK_tX58WsqNgUMHNRIP01LhdK2MDxks5q_2WagaJpZM4K3Isa
.

rpless · 2016-11-19T21:50:54Z

Can do for Spray, but it looks like I gave you the wrong docs for Play Json. We're currently on 2.3.x which does not provide a parse with InputStream. So we can either bump the Play version or System.arraycopy.

vkostyukov · 2016-11-19T22:04:09Z

Arraycopy works for me. Let's update it later.

On Sat, Nov 19, 2016 at 1:50 PM Ryan Plessner notifications@github.com
wrote:

Can do for Spray, but it looks like I gave you the wrong docs for Play
Json. We're currently on 2.3.x
https://www.playframework.com/documentation/2.3.x/api/scala/index.html#play.api.libs.json.Json%24
which does not provide a parse with InputStream. So we can either bump the
Play version or System.arraycopy.

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#671 (comment), or mute
the thread
https://github.com/notifications/unsubscribe-auth/ABDDK22ipWPepiBj6q_IqnkB5M6wpEX7ks5q_28_gaJpZM4K3Isa
.

vkostyukov · 2016-11-20T05:07:08Z

playjson/src/main/scala/io/finch/playjson/package.scala

-    Decode.json((b, cs) => Try(Json.parse(BufText.extract(b, cs)).as[A]))
+    Decode.json { (b, cs) =>
+      val buf = ChannelBufferBuf.Owned.extract(b)
+      if (buf.hasArray) {


Sorry to be a burden, but can we extract that into a function so we can reuse it?

Also, let's check that if array.length equals readableBytes we can skip copying.

Its not a burden. Any particular place we should extract it to? I was thinking io.finch.Decode.

Or maybe the internal package object?

Yeah, let's make it part of internal (whatever format you prefer).

…s conversions in most JSON Decoders.

vkostyukov · 2016-11-21T01:31:30Z

Thanks again @rpless! Merging this.

rpless force-pushed the remove-json-string-conversions branch from 7f464cf to 5194731 Compare November 19, 2016 13:19

vkostyukov requested changes Nov 19, 2016

View reviewed changes

rpless force-pushed the remove-json-string-conversions branch from 5194731 to b0bc72a Compare November 19, 2016 17:49

vkostyukov approved these changes Nov 19, 2016

View reviewed changes

clhodapp reviewed Nov 19, 2016

View reviewed changes

rpless force-pushed the remove-json-string-conversions branch from b0bc72a to c7288c6 Compare November 19, 2016 19:23

vkostyukov requested changes Nov 19, 2016

View reviewed changes

rpless force-pushed the remove-json-string-conversions branch from c7288c6 to 75768b7 Compare November 19, 2016 20:24

rpless force-pushed the remove-json-string-conversions branch from 75768b7 to 582b59a Compare November 19, 2016 22:29

vkostyukov reviewed Nov 20, 2016

View reviewed changes

vkostyukov mentioned this pull request Nov 20, 2016

Introduce decodeByteBuffer, decodeFile, decodeByteArray circe/circe#461

Merged

Replace Buf -> String -> Case Class conversions with Buf -> Case Clas…

a5335d1

…s conversions in most JSON Decoders.

rpless force-pushed the remove-json-string-conversions branch from 582b59a to a5335d1 Compare November 21, 2016 01:10

vkostyukov approved these changes Nov 21, 2016

View reviewed changes

vkostyukov merged commit 80f0523 into finagle:master Nov 21, 2016

vkostyukov mentioned this pull request Nov 21, 2016

Eliminate meaningless to-string conversion #511

Closed

rpless deleted the remove-json-string-conversions branch November 21, 2016 01:33

Remove String conversion in decoders. #671

Remove String conversion in decoders. #671

Conversation

rpless commented Nov 19, 2016

codecov-io commented Nov 19, 2016 • edited Loading

Current coverage is 74.34% (diff: 76.00%)

vkostyukov commented Nov 19, 2016

vkostyukov commented Nov 19, 2016

rpless commented Nov 19, 2016

vkostyukov commented Nov 19, 2016

vkostyukov left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

rpless commented Nov 19, 2016

Without String Conversion (this PR)

With String Conversion (Current Finch)

vkostyukov commented Nov 19, 2016

Choose a reason for hiding this comment

rpless Nov 19, 2016 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

vkostyukov commented Nov 19, 2016

@rpless commented on this pull request.

rpless commented Nov 19, 2016

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

rpless commented Nov 19, 2016

vkostyukov commented Nov 19, 2016

rpless commented Nov 19, 2016

vkostyukov commented Nov 19, 2016

rpless commented Nov 19, 2016

vkostyukov commented Nov 19, 2016

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

vkostyukov commented Nov 21, 2016

codecov-io commented Nov 19, 2016 •

edited

Loading

rpless Nov 19, 2016 •

edited

Loading