A Haskell interface over the simdjson C++ library for decoding JSON documents. Hermes, messenger of the gods, was the maternal great-grandfather of Jason, son of Aeson.
This library exposes functions that can be used to write decoders for JSON documents using the simdjson On Demand API. From the simdjson On Demand design documentation:
Good applications for the On Demand API might be:
You are working from pre-existing large JSON files that have been vetted. You expect them to be well formed according to a known JSON dialect and to have a consistent layout. For example, you might be doing biomedical research or machine learning on top of static data dumps in JSON.
Both the generation and the consumption of JSON data is within your system. Your team controls both the software that produces the JSON and the software the parses it, your team knows and control the hardware. Thus you can fully test your system.
You are working with stable JSON APIs which have a consistent layout and JSON dialect.
With this in mind, Data.Hermes
parsers can decode Haskell types faster than traditional Data.Aeson.FromJSON
instances, especially in cases where you only need to decode a subset of the document. This is because Data.Aeson.FromJSON
converts the entire document into a Data.Aeson.Value
, which means memory usage increases linearly with the input size. The simdjson::ondemand
API does not have this constraint because it iterates over the JSON string in memory without constructing an intermediate tree. This means decoders are truly lazy and you only pay for what you use.
Hermes requires the entire document in memory. For an incremental JSON parser that supports streaming, see json-stream.
This library does not offer a Haskell API over the entire simdjson On Demand API. It currently binds only to what is needed for defining and running a Decoder
. You can see the tests and benchmarks for example usage. simdjson::ondemand
exceptions will be caught and re-thrown with enough information to troubleshoot. In the worst case you may run into a segmentation fault that is not caught, which you are encouraged to report as a bug.
import qualified Data.ByteString as BS
import qualified Data.Hermes as H
personDecoder :: H.Decoder Person
personDecoder = H.object $
Person
<$> H.atKey "_id" H.text
<*> H.atKey "index" H.int
<*> H.atKey "guid" H.text
<*> H.atKey "isActive" H.bool
<*> H.atKey "balance" H.text
<*> H.atKey "picture" (H.nullable H.text)
<*> H.atKey "latitude" H.scientific
-- Decode a strict ByteString.
decodePersons :: BS.ByteString -> Either H.HermesException [Person]
decodePersons = H.decodeEither $ H.list personDecoder
While it is not recommended to use hermes if you need the full DOM, we still provide a performant interface to decode aeson Value
s. See an example of this in the hermes-aeson
subpackage. You could use hermes to selectively decode aeson Value
s on demand, for example:
> decodeEither (atPointer "/statuses/99/user/screen_name" hValueToAeson) twitter
Right (String "2no38mae")
When decoding fails for a known reason, you will get a Left HermesException
indicating if the error came from simdjson
or from an internal hermes
call.
> decodeEither (object . atKey "hello" $ list text) "{ \"hello\": [\"world\", false] }"
Left (SIMDException (DocumentError {path = "/hello/1", errorMsg = "Error while getting value of type text. INCORRECT_TYPE: The JSON element does not have the requested type."}))
We benchmark the following operations using both hermes-json
and aeson
strict ByteString decoders:
- Decode a small array of 3-element arrays of doubles
- Full decoding of a large-ish (12 MB) JSON array of Person objects
- Partial decoding of Twitter status objects to highlight the on-demand benefits
- Decoding entire documents into
Data.Aeson.Value
Please be aware that GHC does not report C-allocated memory. simdjson does actually allocate more memory than appears here, but we still strive to keep our Haskell memory footprint as small as possible.
- GHC 9.8.2 w/ -O1
- aeson-2.2 with text > 2.0
- Apple M1 Pro
Name | Mean (ps) | 2*Stdev (ps) | Allocated | Copied | Peak Memory |
---|---|---|---|---|---|
All.Hermes Arrays | 1176636718 | 112205192 | 4021162 | 42064 | 94371840 |
All.Aeson Arrays | 17213496875 | 475332150 | 71210605 | 1894420 | 94371840 |
All.Hermes Persons | 44490500000 | 2894869504 | 128146854 | 23270453 | 134217728 |
All.Aeson Persons | 129197600000 | 10025634382 | 338386171 | 119316159 | 227540992 |
All.Hermes Partial Twitter | 280497070 | 13556162 | 276618 | 281 | 227540992 |
All.Aeson Partial Twitter | 2791673437 | 155280944 | 11964607 | 187669 | 227540992 |
All.JsonStream Partial Twitter | 2446430468 | 157827966 | 15090720 | 13248 | 227540992 |
All.Hermes Persons (Aeson Value) | 115947050000 | 8397106074 | 276664184 | 105937104 | 227540992 |
All.Aeson Persons (Aeson Value) | 104844225000 | 3647809494 | 269685425 | 96421514 | 250609664 |
All.Hermes Twitter (Aeson Value) | 2933007812 | 223414634 | 11107284 | 189841 | 250609664 |
All.Aeson Twitter (Aeson Value) | 2863515625 | 212391518 | 11652717 | 187635 | 250609664 |
- Decode to
Text
instead ofString
wherever possible! - Decode to
Int
orDouble
instead ofScientific
if you can. - Decode your object fields in order. If encoding with
aeson
, you can leveragetoEncoding
to enforce ordering.
If you need to decode in tight loops or long-running processes (like a server), consider using the withHermesEnv/mkHermesEnv
and parseByteString
functions instead of decodeEither
. This ensures the simdjson instances are not re-created on each decode. See the simdjson performance docs for more info. Please ensure that you use one HermesEnv
per thread, as simdjson is single-threaded by default.
Because the On Demand API in simdjson uses a forward-only iterator (except for object fields), it is possible to introduce unsafe iteration. Hermes tries to prevent this as much as possible with the type system.
The On Demand API does not validate the entire document upon creating the iterator (besides UTF-8 validation and basic well-formed checks). It is possible to parse an invalid JSON document but not realize it until later. If you need the entire document to be validated up front then a DOM parser is a better fit for you.
The On Demand approach is less safe than DOM: we only validate the components of the JSON document that are used and it is possible to begin ingesting an invalid document only to find out later that the document is invalid. Are you fine ingesting a large JSON document that starts with well formed JSON but ends with invalid JSON content?
Other limitations inherited from simdjson:
- Cannot decode scalar documents, e.g. a single string, number, boolean, or null as a JSON document.
- 4GB is the maximum document size that simdjson supports.
Per the simdjson
documentation:
A recent compiler (LLVM clang6 or better, GNU GCC 7.4 or better, Xcode 11 or better) on a 64-bit (PPC, ARM or x64 Intel/AMD) POSIX systems such as macOS, freeBSD or Linux. We require that the compiler supports the C++11 standard or better.
However, this library relies on std::string_view
without a shim, so C++17 or later is required.
The native_comp
cabal flag enables passing -march=native
to the C++ compiler.
Passing -march=native to the compiler may make On Demand faster by allowing it to use optimizations specific to your machine. You cannot do this, however, if you are compiling code that might be run on less advanced machines. That is, be mindful that when compiling with the -march=native flag, the resulting binary will run on the current system but may not run on other systems (e.g., on an old processor).
If you are compiling on an ARM or POWER system, you do not need to be concerned with CPU selection during compilation. The -march=native flag is useful for best performance on x64 (e.g., Intel) systems but it is generally unsupported on some platforms such as ARM (aarch64) or POWER.