Encodes and decodes Thrift binary protocol and JavaScript object models
declared in a Thrift IDL.
This is an alternative approach to using code generated from Thrift IDL source
files with the thrift
compiler.
ThriftRW supports encoding and decoding the protocol with or without Thrift IDL sources. Without sources, it is still possible to read and write the Thrift binary protocol in all of its structure using numbered fields for structs and not discerning the nuanced differences between binary and string, list and set, or number and enum.
With a Thrift IDL, ThriftRW is able to perform certain optimizations. ThriftRW constructs model instances directly from the contents of a buffer instead of creating an intermediate anonymous structural model. As a consequence, it is able to quickly skip any unrecognized fields. ThriftRW can also use the same struct constructor for every instance of a struct, which should yield run-time performance benefits for V8 due to hidden classes (not yet verified).
The scope of this project may eventually cover alternate Thrift binary encodings including the compact binary protocol.
This project makes extensive use of bufrw for reading and writing binary protocols, a component shared by tchannel-node, with which this library works in concert.
Browser compatible, see Browser compatibility.
ThriftRW provides a Thrift constructor that models all of the services, functions, and types expressed in a ThriftIDL source file.
var fs = require('fs');
var path = require('path');
var Thrift = require('thriftrw').Thrift;
var source = fs.readFileSync(path.join(__dirname, 'meta.thrift'), 'utf-8');
var thrift = new Thrift({
source: source,
strict: true,
allowOptionalArguments: true,
defaultAsUndefined: false
});
Consider meta.thrift
struct HealthStatus {
1: required bool ok
2: optional string message
}
service Meta {
HealthStatus health()
string thriftIDL()
}
The most common usage of a Thrift instance is to get the argument and result
types of a function and use those types to read to or write from a buffer.
The getType
and getTypeResult
functions retrieve such models based on their name.
For example, the arguments struct for the health endpoint would be Meta::health_args
and its result struct would be Meta::health_result
.
var MetaHealthArgs = thrift.getType('Meta::health_args');
var MetaHealthResult = thrift.getType('Meta::health_result');
The getType
method will return the struct or throw an exception if one does not exist.
This is the appropriate method if getType
should throw an error due to a
programmer's mistake, where the Thrift source in question is checked into the same project
and the method name is a literal.
However, exceptions being an undesirable outcome for bad data from a user,
getTypeResult
returns a result object with either err
or value
properties set.
var res = thrift.getTypeResult(methodName + '_args');
if (res.err) {
return callback(res.err);
}
var ArgsStruct = res.value;
The struct can be written or read to or from a buffer using the struct's rw
instance.
A RW (reader/writer) implements byteLength(value)
, writeInto(value, buffer, offset)
, readFrom(buffer, offset)
, each of which return a result object as
specified by bufrw.
The value may be JSON, a POJO (plain old JavaScript object), or any instances
of the Thrift model, like new thrift.Health({ok: true})
.
A struct can also be encoded or decoded using the ThriftStruct's own
toBuffer(value)
, fromBuffer(buffer)
, toBufferResult(value)
, and
fromBufferResult(buffer)
methods.
Those with Result
in their name return Result
instances instead of throwing
or returning the buffer or value.
TChannelASThrift employs this interface on your behalf, leaving the application author to take arguments and call back with a result when using the Thrift argument scheme and a given Thrift IDL.
To avoid hazards, every Thrift method should accept a single required argument, numbered 1.
service Service {
void method(1: required MethodArgument arg)
}
struct MethodArgument {
// ...
}
Protobuf gets arguments right. A method receives a single, required message struct. This has very clear semantics in every language.
Thrift methods receive an arguments struct. This struct is special and treated uniquely in each language. Unlike other structs, the field numbers must be consecutive and map to positional arguments in some languages including Go and JavaScript. The argument struct may not even be a named struct.
With the code generated for various languages by Apache Thrift, fields of an argument struct are always optional. The code generator will warn you that marking a field as optional is redundant. If you mark an argument as required, that mark will be silently ignored.
Each language has its own treatment on how to deal with a missing, optional field. In Go, a missing value will be represented as the zero value for its type, effectively having an implicit default value. There is no way to distinguish missing from the zero value. In other languages, it's necessary to check for a missing value (null or None) to avoid exceptions.
The initial version of ThriftRW Node sought to avoid this problem by making all argument fields required, with no recourse for making them optional. In this world, it is not possible to add an argument to an existing method. It is necessary to create a new method. This would be acceptable, except that none of the other ThriftRW implementations were on board with enforcing this.
Consequently, the ThriftRW-node Thrift model constructor accepts an option,
allowOptionalArguments
, which brings Node.js into alignment with ThriftRW for
Python and Go.
All arguments are optional by default, but can be marked required.
This flag is recommended for all future work.
To enable this flag, it is necessary to review all existing methods. For methods that were only previously implemented in Node.js, it is sufficient to mark the arguments as required. This is backward compatible only because they were previously implicitly required. ThriftRW will continue denying requests with the missing arguments. You can then add additional, optional arguments to existing methods, with logic branching on whether the arguments are provided.
For methods that were handled in other languages as well, it is necessary to add run-time checks for null arguments since changing a field from optional to required would be backward incompatible.
Default values are inferred when reading a struct with a missing field, typically by inbound request handlers in servers.
If a client uses the provided constructor for the struct, the constructor will
populate the default value as well, so in that sense clients also infer the
default value. The default value used is the null object unless the defaultAsUndefined
flag is set in the user-defined options. In such a case, undefined will be used, instead.
However, callers are not obliged to use the constructor for the struct. In that case, if the value is missing, ThriftRW does not write it to the wire and instead depends on the server to infer the default.
ThriftRW libraries would be within their rights to detect that a given value matches the default and leave it off the byte buffer when writing. ThriftRW does not yet attempt this because there is no demonstrated need, it could be slow for compound objects, and it could lead to strange behavior if the client and server disagree about the default.
ThriftRW provides T
prefixed types for encoding and decoding the wire
protocol without Thrift IDL.
This is useful for applications like [TCap] that do not necessarily have a
Thrift IDL file on hand to make sense of structs by field numbers alone.
The following example illustrates reading and writing a struct.
ThriftRW exports TStruct
, TList
, and TMap
.
TStructs serve for arguments, results, and exceptions.
TLists also serve for sets.
var thriftrw = require("thriftrw");
var bufrw = require('bufrw');
var struct = new thriftrw.TStruct();
struct.fields.push(
new thriftrw.TField(thriftrw.TYPE.STRING, 1, new Buffer('hello')
);
var buf = bufrw.toBuffer(thriftrw.TStructRW, struct);
console.log('created a binary buffer of thrift encoded struct', buf);
var struct2 = bufrw.fromBuffer(thriftrw.TStructRW, buf);
console.log('created a TStruct from a binary buffer', struct2);
Readers and Writers, or RWs, are objects that capture the methods necessary to read and write a particular value to and from a buffer. RW objects implement:
-
toBuffer(value):Result{err, buffer}
-
fromBuffer(buffer, offset=0):Result{err, value}
-
readFrom(buffer, offset):ReadResult{err, offset, value}
-
writeInto(value, buffer, offset):WriteResult{err, offset, value}
-
byteLength(value):LengthResult{err, length}
-
poolReadFrom(result, buffer, offset):ReadResult{err, offset, value}
-
poolWriteInto(result, value, buffer, offset):WriteResult{err, offset, value}
-
poolByteLength(result, value):LengthResult{err, length}
All of these methods return result objects, which will either have their err
populated with an error, or other properties for the success value.
The pool* methods also accept a result object for reuse.
That result object must implement reset(err, ...properties)
.
The RW is the both the building block and the public interface of ThriftRW.
Thrift has internal models for modules, services, types, and values.
These models are indexed by name on each Thrift module in the models
object.
Each of these models has a "surface", which is a constructor for types, values
for constants, an index of functions for services, and the model itself for
modules.
Modules expose their names on their own object only if they start with a
non-lower-case character, to avoid name collisions with other properties and
methods, but are always also available through the index for their type class,
like thrift.consts.PI
.
The Meta service is simply thrift.Meta
.
The object is an object mapping functions by name, so thrift.Meta.health
is
the interface for accessing the args
and result
structs for the
Meta::health
function.
var args = new thrift.Meta.health.Arguments({ok: true});
var result = new thrift.Meta.health.Result({success: null});
The Thrift instance also has a services
property containing the actual
ThriftService instance indexed by service name.
ThriftService instances are not expected to be useful outside the process of
compiling and linking the Thrift IDL, but you can use the services object to
check for the existence of a service by its name.
Apache Thrift over HTTP envelopes arguments and result structs with a message that contains a sequence number, the function name, a message type, and followed by the message body.
Thrift over TChannel does not use Thrift envelopes. TChannel does not need or use the Thrift sequence number, since it abstracts multiplexing with its own transport frame identifiers. The function name is expressed as Service::function in TChannel's arg1. The message type corresponds to the TChannel frame type. The message body corresponds to TChannel's arg3.
Since HTTP/1.1 only allows one outstanding request and response on the wire at any time, and HTTP/1.1 with the ill-fated "pipeline" mode allows requests and responses to be sent FIFO but forbids out-of-order responses, the sequence number is useless. HTTP/2.0 does support multiplexing, but does not require sequence identifiers in the request or response body to do so. Despite this folly, HTTP request handlers are obliged to use envelopes and match the number of a response to the number of a request, but callers are not obliged to choose sensible sequence numbers.
By default, services dispatch on the given function name. Services that implement multiple Thrift services have the option of requiring the Thrift service name and function name in the message envelope, with an arbitrary delimiter, typically a single colon, e.g., "Service:function". Over TChannel, this corresponds to the frame's arg1 and is always "Service::function", the service name and function name delimited by two colons.
The message type is an enumeration of one of 'CALL', 'ONEWAY', 'RESULT', or
'EXCEPTION'.
'CALL' and 'ONEWAY' are appropriate for requests, where 'ONEWAY' is for
functions with the oneway
Thrift IDL modifier keyword, implying they expect
no response.
'RESULT' is for both Thrift success and exception cases.
For exceptions that cannot be captured by the Thrift result struct, the
'EXCEPTION' message type implies that the body is a Thrift exception struct of
the following shape instead of the result struct from the IDL.
enum ExceptionType {
UNKNOWN = 0
UNKNOWN_METHOD = 1
INVALID_MESSAGE_TYPE = 2
WRONG_METHOD_NAME = 3
BAD_SEQUENCE_ID = 4
MISSING_RESULT = 5
INTERNAL_ERROR = 6
PROTOCOL_ERROR = 7
INVALID_TRANSFORM = 8
INVALID_PROTOCOL = 9
UNSUPPORTED_CLIENT_TYPE = 10
}
struct TApplicationException {
1: optional string message
2: optional ExceptionType type
}
These cases are expressed in TChannel instead with an Error frame with the corresponding TChannel transport error code and an arbitrary message.
The base message envelope constructor is available as the thrift.Message
,
that is also a copy constructor.
Messages have the following properties:
version
defaults to 0. Version 1 has a more elaborate message header that expresses the message envelope version number and provides no additional value. There may or may never be another message envelope version.type
is one of 'CALL', 'ONEWAY', 'RESULT', or 'EXCEPTION'id
is a sequence number. Really, it is any number. If you are a client, you can pick. If you are a server, you must respect the client's choice and send back the same id in the response message.body
is either your result object, or aTApplicationException
.
ThriftRW provides reader/writer objects for the arguments and result message
envelopes on the function model.
The function model is addressable as thrift.{Service}.{function}
.
argumentsMessageRW
resultMessageRW
ThriftStruct models can be constructed with an optional object of specified
properties.
ThriftRW exposes the constructor for a struct declaration by name on the thrift
instance, as in thrift.Health
for the meta.thrift example.
var HealthStruct = thrift.Health;
var result = new ResultStruct({
success: new HealthStruct({ok: true })
})
Unspecified properties will default to null (undefined if defaultAsUndefined flag was set) or an instance of the default value specified in the Thrift IDL. Nested structs can be expressed with JSON or POJO equivalents. The constructors perform no validation. Invalid objects are revealed only in the attempt to write one to a buffer.
var result = new ResultStruct({success: {ok: true}});
Each constructor has a rw
property that reveals the reader/writer instance.
Each constructor also hosts toBuffer(value)
, fromBuffer(buffer)
,
toBufferResult(value)
, and fromBufferResult(buffer)
.
Structs are indexed by name on the thrift.structs
object and aliased on the
thrift object if their name does not start with a lower-case letter.
ThriftException extends ThriftStruct. Exceptions are modeled as structs on the wire, and are modeled as JavaScript exceptions instead of regular objects. As such they have a stack trace. ThriftRW exposes the exception constructor by name on the thrift instance.
exception Pebcak {
1: required string message
2: required string keyboardName
3: required string chairName
}
var error = new thrift.Pebcak({
message: 'Problem exists between chair and keyboard',
chairName: 'Hengroen',
keyboardName: 'Excalibur'
});
Exceptions are indexed by name on the thrift.exceptions
object and aliased on
the thrift object if their name does not start with a lower-case letter.
Types, services, and constants defined in different Thrift files may be referenced by using include statements with paths relative to the current .thrift file. The paths must be in the form ./foo.thrift, ./foo/bar.thrift, ../baz.thrift, and so on.
Included modules will automatically be interpreted along with the module that included them, and they will be made available in the generated module with the base name of the included file.
For example, given:
// shared/types.thrift
struct UUID {
1: required i64 high
2: required i64 low
}
And:
// services/user.thrift
include "../shared/types.thrift"
struct User {
1: required types.UUID uuid
}
You can do the following
var path = require('path');
var Thrift = require('thriftrw').Thrift;
var service = new Thrift({
entryPoint: path.resolve(__dirname, 'services/user.thrift'),
allowFilesystemAccess: true
});
var userUuid = service.modules.types.UUID({
high: ...,
low: ...
});
var user = service.User({
uuid: userUuid
});
thriftrw-node contains experimental include-as support. This lets you include a file aliasing it to a different name than that of the file. This feature is hidden behind a flag and should not be used until support for include-as aliasing is supported by other language implementations of thriftrw.
Unaliased include:
include "../shared/Types.thrift"
Aliased include:
include Types "../shared/types.thrift"
To enable this you need to create a Thrift
with the allowIncludeAlias
option set to true. e.g.
var thrift = new Thrift({
thriftFile: /* path to main thrift file */,
allowIncludeAlias: true,
allowFilesystemAccess: true
});
The surface of any Thrift object will expose definitions as top level properties on an object if the identifier for that definition is PascalCased. In the include example above this can be seen. The filename for the types.thrift
file is not PascalCase so it needed to be reached via the modules property. If the filename had instead been PascalCased, like Types.thrift
or imported as Types
, that module could have been accessed directly via service.Types.UUID
;`
ThriftUnion also extends ThriftStruct.
Unions are alike to structs except that fields must not be marked as optional
or required
, cannot have a default value, and exactly one must be defined on
an instance.
As with other types, validation of a union only occurs when it is read or
written.
union CoinToss {
1: Obverse heads
2: Reverse tails
}
struct Obverse {
1: required string portrait;
2: required i32 year;
}
struct Reverse {
1: required string structure;
2: optional string motto;
}
var coinToss = new thrift.CoinToss({
head: thrift.Obverse({
portrait: 'TK',
year: 2010
})
})
Unions are indexed by name on the thrift.unions
object and aliased on the
thrift object if their name does not start with a lower-case letter.
Thrift enumerations are surfaced as an object mapping names to strings on the thrift instance.
enum CoinToss {
tails = 0
heads = 1
}
var result = Math.random() < 0.5 ?
thrift.CoinToss.heads :
thrift.CoinToss.tails;
// result will be either "tails" or "heads"
ThriftRW hides the detail of which numeric value an enumeration will have on the wire and allows an enumeration to be represented in JSON by name. This is a deliberate departure from the norms for Thrift in other languages to take advantage of the dynamic nature of JavaScript, permit duck typing, enforce loose coupling between application code and wire protocol, and make Thrift expressible as JSON on the command line with TCurl
Enums are indexed by name on the thrift.unions
object and aliased on the
thrift object if their name does not start with a lower-case letter.
Thrift constants are surfaced as properties of the thrift instance.
const PI = 3
const TAU = 6
var PI = thrift.PI;
var TAU = thrift.TAU;
Consts are indexed by name on the thrift.consts
object and aliased on the
thrift object if their name does not start with a lower-case letter.
ThriftRW operates in a "strict" mode by default that imposes additional constraints on Thrift to ensure cross language interoperability between JavaScript, Python, Go, and Java, the four Thrift bindings favored at Uber.
In strict mode, structs must explicitly declare every field to be either required, optional, or have a default value.
ThriftRW supports forward references and self-referential types (like trees) without special demarcation. ThriftRW also supports forward references in default values and constants.
ThriftRW respects certain annotations specific to the treatment of various types in JavaScript.
There is an emerging standard for the representation of sets in JavaScript. Until that standard becomes pervasive, sets are traditionally represented either as an array with externally enforced constraints, or as long as the values are scalar like numbers and strings, as an Object with values for keys.
By default, ThriftRW encodes and decodes sets as arrays. ThriftRW does nothing to enforce uniqueness.
Additionally, the js.type
annotation can be set to "object" if the value type
is a string, i16, or i32.
ThriftRW will ensure that the keys are coerced to and from Number for numeric
value types.
typedef set<string> (js.type = 'object') StringSet
typedef set<i32> (js.type = 'object') I32Set
There is also an emerging standard for the representation of maps in JavaScript. However, until support is more pervasive, ThriftRW supports representing maps as either an object with scalar key types or an entries array for all others.
Maps are represented as objects by default. With the js.type
annotation set
to "entries", the map will be encoded and decoded with an array of [key, value]
duples.
typedef map<string, string> Dictionary
typedef map<Struct, Struct> (js.type = 'entries') StructsByStructs
JavaScript numbers lack sufficient precision to represent all possible 64 bit integers. ThriftRW decodes 64 bit integers into a Buffer by default, and can coerce various types down to i64 for purposes of expressing them as JSON or plain JavaScript objects.
- A number up to the maximum integer precision available in JavaScript.
- A
{hi, lo}
or{high, low}
pair of 32 bit precision integers. - A 16 digit hexadecimal string, like
0102030405060708
. - An array of 8 byte values. Ranges are not checked, but coerced.
- An 8 byte buffer.
ThriftRW supports a type annotation for i64 that reifies i64 as an instance of
Long, an object that models a 64 bit number as a {high, low}
pair of 32
bit numbers and implements methods for common mathematical operators.
typedef i64 (js.type = 'Long') long
Please use the Long package for operating with such numbers. A backward-incompatible release may make this form the default for reading.
With the Long type annotation, ThriftRW performs the following conversions on your behalf.
var Long = require('long');
// From a buffer to a long
var buf = new Buffer("0102030405060708", "hex");
var num = Long.fromBits(buf.readInt32BE(4, true), buf.readInt32BE(0, true));
// From a long to a buffer
var buf = new Buffer(8);
buf.writeUInt32BE(num.high, 0, 4, true);
buf.writeUInt32BE(num.low, 4, 4, true);
The convention for expressing timestamps at Uber is to use an i64 for
UTC milliseconds since the UNIX epoch.
Previously, some services were found spending an unreasonable amount of time
parsing timestamp strings.
ThriftRW supports a (js.type = 'Date')
annotation on i64.
typedef i64 (js.type = 'Date') timestamp
This will reify timestamps as a JavaScript Date and ThriftRW will accept an
ISO-8601 timestamp or milliseconds since the epoch (as returned by
Date.now()
) in place of a Date, as well as the equivalent in buffers or
arrays of bytes.
For TCurl, this means that you can both read and write ISO-8601 for timestamps.
You can use the alternative asynchronous constructor Thrift.load()
to load sources from a custom asynchronous function. The function provided will be passed file names and must call the callback function with the source associated with the file name. The callback function provided as second argument to Thrift.load
will be called when the model is fully loaded or an error occured. This way, you can fetch thrift sources from a server for example. Here is an example:
var request = require('request');
var Thrift = require('thriftrw').Thrift;
function fetchSource(filename, cb) {
request('https://mysite.com/thrift/' + filename, function (err, res, body) {
cb(err, body);
});
}
// You can pass the `fs` Node.js module like so {fs: require('fs')}
var options = {fs: {readFile: fetchSource}};
Thrift.load(options, function (err, thrift) {
console.log(err, thrift);
});
The browser/dist/bundle.js
file is a browserify bundle of thriftrw compatible with browsers. If you require('thriftrw')
in the context of a browser build, your build tool will likely use the browser
field in package.json
and use
the bundle instead of the node library (e.g. browserify, webpack). If it doesn't, you can probably make it do so.
The only thing that won't work in a browser is file system access. To go around this limitation, you have several aternatives:
- Load sources using a custom function, which can be asynchronous. See Non filesystem and asynchronous source loading.
- Use the
source
option as argument tonew Thrift()
- Set all required files in the
idls
option as argument tonew Thrift()
, requiresentryPoint
to be specified, and thesource
not to be specified - Pass an object with a custom
readFileSync
function as thefs
argument tonew Thrift()
, requiresallowFilesystemAccess
to be falsy, andsource
orentryPoint
to be specified
thriftrw-python is the sister library for Python.
npm install thriftrw
npm test
npm run add-licence
This will add the licence headers.npm run build-browser
This will build the test and release browserify bundles.npm run cover
This runs the tests with code coveragenpm run lint
This will run the linter on your codenpm test
This will run the tests.npm run test-browser
This will run the tests using the browserify bundle in Headless Chrome.npm run trace
This will run your tests in tracing mode.npm run travis
This is run by travis.CI to run your testsnpm run view-cover
This will show code coverage in a browser
- Lei Zhao @leizha
- Kris Kowal @kriskowal
- Andrew de Andrade @malandrew