-
|
For an example when parsing json string to something like binary json, it's really better than ondemand. And I consider other usecases exists, althouth for most usecases dom/ondemand better alternative. Some example: |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment · 7 replies
-
|
it's really better than ondemand. We designed On-Demand specifically as a replacement for event-based parsing. We had, for a time, an event-based API but we removed it. It is not faster and I feel that for most people, it is harder to use. If you find On-Demand difficult to use for some use cases, please raise specific issues, we will try to address it. |
Beta Was this translation helpful? Give feedback.
-
|
with ondemand such use case needs to be recursive :( Is any option how to use ondemand without recursion? |
Beta Was this translation helpful? Give feedback.
-
This code looks very good to me: https://github.com/ydb-platform/ydb/blob/74df9273b222dc253487508e6f4602237f4a7c11/ydb/library/binary_json/write.cpp#L553 ❤️ We get pretty decent results with recursion and the code is relatively elegant. The one annoying caveat is that GCC and LLVM differ in how they handle recursion in practice, so a bit of care is needed if one wants to get good performance. But that's relatively minor. Please see... https://github.com/simdjson/simdjson/blob/master/benchmark/json2msgpack/simdjson_ondemand.h Depending on your compiler, you might get better result when passing by value instead of by reference.
You can handle the stack yourself if you prefer. The stack is either made of an object and its iterator, or an array and its iterator. But that's going to be annoying to implement. However, in simdjson, we already have a JSON to binary conversion routine that runs very fast. It is our DOM builder. And it is done by an event-like system. Please see... https://github.com/simdjson/simdjson/blob/master/src/generic/stage2/tape_builder.h For example, we have... The key ingredient is the 'tape' which receives the values. Obviously, this would work with any "tape"... We would need to go back in the git tree, but this used to be exposed in some way (but we never documented it). I'd be concerned about opening this up to everyone because very few people should program this way. It is error prone. However, if you are willing to help, we could make it usable from people like ydb. It is not very hard because we already use it internally. It is mostly a matter of plumbing. Let me know if this is something you'd like to help with!!! |
Beta Was this translation helpful? Give feedback.
-
|
tape builder looks really interesting, I will try to play with this code a little and benchmark it with ondemand approach, thanks I also thought about we have stack in binary json builder and it can have additional payload to use it instead of recursion |
Beta Was this translation helpful? Give feedback.
-
|
@MBkkt To be honest, I'd be interested in us having a routine that is suitable to write very fast JSON -> something else conversions. If we focus on this one use case, I think we can write something very efficient and very nice. |
Beta Was this translation helpful? Give feedback.
-
|
Sounds awesome, maybe later I will try to make type_builder |
Beta Was this translation helpful? Give feedback.
-
|
Ofc if someone will do it before me, I will be happy :) |
Beta Was this translation helpful? Give feedback.
@MBkkt
This code looks very good to me: https://github.com/ydb-platform/ydb/blob/74df9273b222dc253487508e6f4602237f4a7c11/ydb/library/binary_json/write.cpp#L553
❤️
We get pretty decent results with recursion and the code is relatively elegant. The one annoying caveat is that GCC and LLVM differ in how they handle recursion in practice, so a bit of care is needed if one wants to get good performance. But that's relatively minor.
Please see...
https://github.com/simdjson/simdjson/blob/master/benchmark/json2msgpack/simdjson_ondemand.h
Depending on your compiler, you might get better result when passing by value instead of by reference.