When dumping to IO, dump directly #538
Closed
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
headius
force-pushed
the
streaming_output
branch
from
2766f66
to
7576f5c
Compare
|
Note that prior to this patch, the script provided in jruby/jruby#6265 required 2-3GB of memory to run. Afterwards it varies between 200-600MB and completes successfully. |
|
For performance, it might help to have a buffered proxy to the underlying IO? That way, you only incur the locking/encoding overhead every buffer block size, vs every substring that gets written? |
|
@segiddins Yeah I hoped that internal buffering in IO would be sufficient for this but that buffer may simply be too small or there's enough overhead with character encoding checks that we lose the benefit. I'd like to play with some other buffering strategies, and some of the larger methods that do many tiny writes could do those rights into a temporary string that gets dumped in one go. |
|
Oh, I did run a profile of this on JRuby and the other cost we have using IO as the buffer is all the locking that's required to keep it thread safe. So if you're writing individual characters, that's a lot of lock acquisition and releasing. Batching up those rights into coarser chunks would make a big difference. |
|
I'm assuming you're not actively working on this, so I'll close. |
|
@byroot I never stopped working on this but never got feedback from any stakeholders. I still believe it should be done. |
|
FWIW we have a customer that needed this, and I assume they still do. Dumping directly to an in-memory buffer is prohibitive for very large json streams. |
|
Yeah, it makes sense, and looking at the C implementation, I think it would be relatively easy to do the C implementation. I'm just trying to keep the repo tidy, if you feel strongly about keeping that draft open feel free to re-open.
Well, as mentioned previously, from my point of view you are the Java implementation maintainer, so if the feature doesn't require to change the public API, feel fee to implement and merge whatever you want. |
|
@byroot I was not a maintainer at the time, so I was not enthusiastic to do much with it. That's why it sat for a year+ without any work being done; if the maintainer was not on board, there would not be much point in me doing it. I also did not hear from a single maintainer of the C extension and I do not have the knowledge of that codebase to make a similar change, so that further dampened my enthusiasm. I will re-open. I believe we should do this for both extensions and json-pure. |
As for the C extension, I can take care of it, but the nice thing about this feature is that AFAICT it doesn't change anything about the public API it's just an internal implementation detail, so IMO we can perfectly merge the Java side of it and see later for the C side. |
Ref: ruby#524 Rather than to buffer everything in memory. Unfortunately Ruby doesn't provide an API to write into and IO without first allocating a string, which is a bit wasteful.
|
@byroot That makes sense. I will return to this and get it working well for the Java extension. If it seems like a useful optimization at that point, we'll proceed with a C impl. |
|
Most of the C implementation is in #686, just need to take the time to polish it. |
Ignoring `CHAR_BITS` > 8 platform, as far as `ch` indexes `escape_table` that is hard-coded as 256 elements. ``` ../../../../src/ext/json/generator/generator.c(121): warning C4333: '>>': right shift by too large amount, data loss ../../../../src/ext/json/generator/generator.c(122): warning C4333: '>>': right shift by too large amount, data loss ../../../../src/ext/json/generator/generator.c(243): warning C4333: '>>': right shift by too large amount, data loss ../../../../src/ext/json/generator/generator.c(244): warning C4333: '>>': right shift by too large amount, data loss ../../../../src/ext/json/generator/generator.c(291): warning C4333: '>>': right shift by too large amount, data loss ../../../../src/ext/json/generator/generator.c(292): warning C4333: '>>': right shift by too large amount, data loss ```
If we assume most string don't contain any escape sequence we can avoid
a lot of costly operations when it holds true.
Before:
```
== Parsing activitypub.json (58160 bytes)
ruby 3.4.0dev (2024-11-06T07:59:09Z precompute-hash-wh.. 7943f98a8a) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
json 884.000 i/100ms
oj 789.000 i/100ms
Oj::Parser 943.000 i/100ms
rapidjson 584.000 i/100ms
Calculating -------------------------------------
json 8.897k (± 1.3%) i/s (112.40 μs/i) - 45.084k in 5.068520s
oj 7.967k (± 1.5%) i/s (125.52 μs/i) - 40.239k in 5.051985s
Oj::Parser 9.564k (± 1.4%) i/s (104.56 μs/i) - 48.093k in 5.029626s
rapidjson 5.947k (± 1.4%) i/s (168.16 μs/i) - 29.784k in 5.009437s
Comparison:
json: 8896.5 i/s
Oj::Parser: 9563.8 i/s - 1.08x faster
oj: 7966.8 i/s - 1.12x slower
rapidjson: 5946.7 i/s - 1.50x slower
== Parsing twitter.json (567916 bytes)
ruby 3.4.0dev (2024-11-06T07:59:09Z precompute-hash-wh.. 7943f98a8a) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
json 83.000 i/100ms
oj 64.000 i/100ms
Oj::Parser 77.000 i/100ms
rapidjson 54.000 i/100ms
Calculating -------------------------------------
json 823.083 (± 1.8%) i/s (1.21 ms/i) - 4.150k in 5.043805s
oj 632.538 (± 1.4%) i/s (1.58 ms/i) - 3.200k in 5.060073s
Oj::Parser 769.122 (± 1.8%) i/s (1.30 ms/i) - 3.850k in 5.007501s
rapidjson 548.494 (± 1.5%) i/s (1.82 ms/i) - 2.754k in 5.022153s
Comparison:
json: 823.1 i/s
Oj::Parser: 769.1 i/s - 1.07x slower
oj: 632.5 i/s - 1.30x slower
rapidjson: 548.5 i/s - 1.50x slower
== Parsing citm_catalog.json (1727030 bytes)
ruby 3.4.0dev (2024-11-06T07:59:09Z precompute-hash-wh.. 7943f98a8a) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
json 41.000 i/100ms
oj 34.000 i/100ms
Oj::Parser 45.000 i/100ms
rapidjson 39.000 i/100ms
Calculating -------------------------------------
json 427.162 (± 1.2%) i/s (2.34 ms/i) - 2.173k in 5.087666s
oj 351.463 (± 2.8%) i/s (2.85 ms/i) - 1.768k in 5.035149s
Oj::Parser 461.849 (± 3.7%) i/s (2.17 ms/i) - 2.340k in 5.074461s
rapidjson 395.155 (± 1.8%) i/s (2.53 ms/i) - 1.989k in 5.034927s
Comparison:
json: 427.2 i/s
Oj::Parser: 461.8 i/s - 1.08x faster
rapidjson: 395.2 i/s - 1.08x slower
oj: 351.5 i/s - 1.22x slower
```
After:
```
== Parsing activitypub.json (58160 bytes)
ruby 3.4.0dev (2024-11-06T07:59:09Z precompute-hash-wh.. 7943f98a8a) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
json 953.000 i/100ms
oj 813.000 i/100ms
Oj::Parser 956.000 i/100ms
rapidjson 563.000 i/100ms
Calculating -------------------------------------
json 9.525k (± 1.2%) i/s (104.98 μs/i) - 47.650k in 5.003252s
oj 8.117k (± 0.5%) i/s (123.20 μs/i) - 40.650k in 5.008283s
Oj::Parser 9.590k (± 3.2%) i/s (104.27 μs/i) - 48.756k in 5.089794s
rapidjson 6.020k (± 0.9%) i/s (166.10 μs/i) - 30.402k in 5.050155s
Comparison:
json: 9525.3 i/s
Oj::Parser: 9590.1 i/s - same-ish: difference falls within error
oj: 8116.7 i/s - 1.17x slower
rapidjson: 6020.5 i/s - 1.58x slower
== Parsing twitter.json (567916 bytes)
ruby 3.4.0dev (2024-11-06T07:59:09Z precompute-hash-wh.. 7943f98a8a) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
json 87.000 i/100ms
oj 64.000 i/100ms
Oj::Parser 75.000 i/100ms
rapidjson 55.000 i/100ms
Calculating -------------------------------------
json 866.563 (± 0.8%) i/s (1.15 ms/i) - 4.350k in 5.020138s
oj 643.567 (± 0.8%) i/s (1.55 ms/i) - 3.264k in 5.072101s
Oj::Parser 777.346 (± 3.5%) i/s (1.29 ms/i) - 3.900k in 5.023933s
rapidjson 557.158 (± 0.7%) i/s (1.79 ms/i) - 2.805k in 5.034731s
Comparison:
json: 866.6 i/s
Oj::Parser: 777.3 i/s - 1.11x slower
oj: 643.6 i/s - 1.35x slower
rapidjson: 557.2 i/s - 1.56x slower
== Parsing citm_catalog.json (1727030 bytes)
ruby 3.4.0dev (2024-11-06T07:59:09Z precompute-hash-wh.. 7943f98a8a) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
json 41.000 i/100ms
oj 35.000 i/100ms
Oj::Parser 40.000 i/100ms
rapidjson 39.000 i/100ms
Calculating -------------------------------------
json 429.216 (± 1.2%) i/s (2.33 ms/i) - 2.173k in 5.063351s
oj 354.755 (± 1.1%) i/s (2.82 ms/i) - 1.785k in 5.032374s
Oj::Parser 465.114 (± 3.7%) i/s (2.15 ms/i) - 2.360k in 5.081634s
rapidjson 387.135 (± 1.3%) i/s (2.58 ms/i) - 1.950k in 5.037787s
Comparison:
json: 429.2 i/s
Oj::Parser: 465.1 i/s - 1.08x faster
rapidjson: 387.1 i/s - 1.11x slower
oj: 354.8 i/s - 1.21x slower
```
`rb_cstr2inum` isn't very fast because it handles tons of
different scenarios, and also require a NULL terminated string
which forces us to copy the number into a secondary buffer.
But since the parser already computed the length, we can much more
cheaply do this with a very simple function as long as the number
is small enough to fit into a native type (`long long`).
If the number is too long, we can fallback to the `rb_cstr2inum`
slowpath.
Before:
```
== Parsing citm_catalog.json (1727030 bytes)
ruby 3.4.0dev (2024-11-06T07:59:09Z precompute-hash-wh.. 7943f98a8a) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
json 40.000 i/100ms
oj 35.000 i/100ms
Oj::Parser 45.000 i/100ms
rapidjson 38.000 i/100ms
Calculating -------------------------------------
json 425.941 (± 1.9%) i/s (2.35 ms/i) - 2.160k in 5.072833s
oj 349.617 (± 1.7%) i/s (2.86 ms/i) - 1.750k in 5.006953s
Oj::Parser 464.767 (± 1.7%) i/s (2.15 ms/i) - 2.340k in 5.036381s
rapidjson 382.413 (± 2.4%) i/s (2.61 ms/i) - 1.938k in 5.070757s
Comparison:
json: 425.9 i/s
Oj::Parser: 464.8 i/s - 1.09x faster
rapidjson: 382.4 i/s - 1.11x slower
oj: 349.6 i/s - 1.22x slower
```
After:
```
== Parsing citm_catalog.json (1727030 bytes)
ruby 3.4.0dev (2024-11-06T07:59:09Z precompute-hash-wh.. 7943f98a8a) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
json 46.000 i/100ms
oj 33.000 i/100ms
Oj::Parser 45.000 i/100ms
rapidjson 39.000 i/100ms
Calculating -------------------------------------
json 462.332 (± 3.2%) i/s (2.16 ms/i) - 2.346k in 5.080504s
oj 351.140 (± 1.1%) i/s (2.85 ms/i) - 1.782k in 5.075616s
Oj::Parser 473.500 (± 1.3%) i/s (2.11 ms/i) - 2.385k in 5.037695s
rapidjson 395.052 (± 3.5%) i/s (2.53 ms/i) - 1.989k in 5.042275s
Comparison:
json: 462.3 i/s
Oj::Parser: 473.5 i/s - same-ish: difference falls within error
rapidjson: 395.1 i/s - 1.17x slower
oj: 351.1 i/s - 1.32x slower
```
And stop benchmarking the Generator::State re-use, as it no longer make a sizeable difference.
Fix: ruby#694 This was lost during the .gemspec merge and not noticed because it was falling back to loading the jars from the stdlib.
This way they're hidden in diffs. It would be good to enforce on CI that the generated files match the source change, however ragel's output isn't consistent across versions and system, so we'll have to rely on changes being noticed by further contributions.
Before this commit, we would try to scan for a float, then if that failed, scan for an integer. But floats and integers have many bytes in common, so we would end up scanning the same bytes multiple times. This patch combines integer and float scanning machines so that we only have to scan bytes once. If the machine finds "float parts", then it executes the "isFloat" transition in the machine, which sets a boolean letting us know that the parser found a float. If we didn't find a float, but we did match, then we know it's an int.
Fix: ruby#697 This way even if `Encoding.default_external` is set to a weird value the document will be parsed just fine.
The documentation state `Oj::Parser.usual` isn't thread safe: https://github.com/ohler55/oj/blob/c70bf4125b546bc7146840b15de36460d42b4dff/ext/oj/parser.c#L1507-L1513 As such we shouldn't benchark it this way, but instantiate a new parser every time. Technically in real world scenarios you could create a pool of parsers and re-use them, but if it's not provided by the gem, I'm not sure we should go out of our way to do it.
Otherwise the likeliness of seeing that key again is really low, and looking up
the cache is just a waste.
Before:
```
== Parsing small hash (65 bytes)
ruby 3.4.0dev (2024-11-13T12:32:57Z fstr-update-callba.. 9b44b455b3) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
json 343.049k i/100ms
oj 213.943k i/100ms
Oj::Parser 31.583k i/100ms
rapidjson 303.433k i/100ms
Calculating -------------------------------------
json 3.704M (± 1.5%) i/s (270.01 ns/i) - 18.525M in 5.003078s
oj 2.200M (± 1.1%) i/s (454.46 ns/i) - 11.125M in 5.056526s
Oj::Parser 285.369k (± 4.8%) i/s (3.50 μs/i) - 1.453M in 5.103866s
rapidjson 3.216M (± 1.6%) i/s (310.95 ns/i) - 16.082M in 5.001973s
Comparison:
json: 3703517.4 i/s
rapidjson: 3215983.0 i/s - 1.15x slower
oj: 2200417.1 i/s - 1.68x slower
Oj::Parser: 285369.1 i/s - 12.98x slower
== Parsing test from oj (258 bytes)
ruby 3.4.0dev (2024-11-13T12:32:57Z fstr-update-callba.. 9b44b455b3) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
json 54.539k i/100ms
oj 41.473k i/100ms
Oj::Parser 24.064k i/100ms
rapidjson 51.466k i/100ms
Calculating -------------------------------------
json 549.386k (± 1.6%) i/s (1.82 μs/i) - 2.781M in 5.064316s
oj 417.003k (± 1.3%) i/s (2.40 μs/i) - 2.115M in 5.073047s
Oj::Parser 226.500k (± 4.7%) i/s (4.42 μs/i) - 1.131M in 5.005466s
rapidjson 526.124k (± 1.0%) i/s (1.90 μs/i) - 2.676M in 5.087176s
Comparison:
json: 549385.6 i/s
rapidjson: 526124.3 i/s - 1.04x slower
oj: 417003.4 i/s - 1.32x slower
Oj::Parser: 226500.4 i/s - 2.43x slower
```
After:
```
== Parsing small hash (65 bytes)
ruby 3.4.0dev (2024-11-13T12:32:57Z fstr-update-callba.. 9b44b455b3) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
json 361.394k i/100ms
oj 217.203k i/100ms
Oj::Parser 28.855k i/100ms
rapidjson 303.404k i/100ms
Calculating -------------------------------------
json 3.859M (± 2.9%) i/s (259.13 ns/i) - 19.515M in 5.061302s
oj 2.191M (± 1.6%) i/s (456.49 ns/i) - 11.077M in 5.058043s
Oj::Parser 315.132k (± 7.1%) i/s (3.17 μs/i) - 1.587M in 5.065707s
rapidjson 3.156M (± 4.0%) i/s (316.88 ns/i) - 15.777M in 5.008949s
Comparison:
json: 3859046.5 i/s
rapidjson: 3155778.5 i/s - 1.22x slower
oj: 2190616.0 i/s - 1.76x slower
Oj::Parser: 315132.4 i/s - 12.25x slower
== Parsing test from oj (258 bytes)
ruby 3.4.0dev (2024-11-13T12:32:57Z fstr-update-callba.. 9b44b455b3) +YJIT +PRISM [arm64-darwin24]
Warming up --------------------------------------
json 55.682k i/100ms
oj 40.343k i/100ms
Oj::Parser 25.119k i/100ms
rapidjson 51.500k i/100ms
Calculating -------------------------------------
json 555.808k (± 1.4%) i/s (1.80 μs/i) - 2.784M in 5.010092s
oj 412.283k (± 1.7%) i/s (2.43 μs/i) - 2.098M in 5.089900s
Oj::Parser 279.306k (±13.3%) i/s (3.58 μs/i) - 1.356M in 5.022079s
rapidjson 517.177k (± 2.7%) i/s (1.93 μs/i) - 2.626M in 5.082352s
Comparison:
json: 555808.3 i/s
rapidjson: 517177.1 i/s - 1.07x slower
oj: 412283.2 i/s - 1.35x slower
Oj::Parser: 279306.5 i/s - 1.99x slower
```
|
@byroot Your approach is pretty close to the new attempt I'm making. I will rebase my work on this PR. |
|
Correction, I will rebase my PR on your PR #686 once I update it for the recent changes. |
When an IO is given, we should try to write directly to it. This patch moves that direction by always doing JSON dumping into an OutputStream, which can be implemented on top of a given IO or by producing a ByteList via a ByteArrayOutputStream.
This connects up the OutputStream-based generator logic to the incoming IO parameter (an IO or nil). Also included here are some small changes to support the new state.rb: * Require the state.rb file at the end of ext init. * Move State#configure to _configure. * Add State#strict? as an alias for strict.
headius
force-pushed
the
streaming_output
branch
from
7576f5c
to
2ebe42c
Compare
|
@byroot Github seems to be confused about how this has been rebased. The two commits I made here are the only two for JRuby, and the others should already be on master somewhere. Perhaps my rebase from your branch has confused it too much. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Note, this PR is a proof-of-concept of direct IO dumping in the json library. It works, but is not as fast as it could be, and there's no CRuby implementation yet.
Json.dump allows you to pass an IO to which the dump output will be sent, but it still buffers the entire output in memory before sending it to the given IO. This leads to issues on JRuby like jruby/jruby#6265 when it tries to create a byte[] that exceeds the maximum size of a signed int (JVM's array size limit).
This commit plumbs the IO all the way through the generation logic so that it can be written to directly without filling a temporary memory buffer first. This allow JRuby to dump object graphs that would normally produce more content than the JVM can hold in a single array, providing a workaround for jruby/jruby#6265.
It is unfortunately a bit slow to dump directly to IO due to the many small writes that all acquire locks and participate in the IO encoding subsystem. A more direct path that can skip some of these pieces could be more competitive with the in-memory version, but functionally it expands the size of graphs that cana be dumped when using JRuby.
See #524