Fetching contributors…
Cannot retrieve contributors at this time
1058 lines (885 sloc) 47.6 KB
% Thrift whitepaper
% Name: thrift.tex
% Authors: Mark Slee (
% Created: 05 March 2007
% You will need a copy of sigplanconf.cls to format this document.
% It is available at <>.
% \conferenceinfo{WXYZ '05}{date, City.}
% \copyrightyear{2007}
% \copyrightdata{[to be supplied]}
% \titlebanner{banner above paper title} % These are ignored unless
% \preprintfooter{short description of paper} % 'preprint' option specified.
\title{Thrift: Scalable Cross-Language Services Implementation}
\authorinfo{Mark Slee, Aditya Agarwal and Marc Kwiatkowski}
{Facebook, 156 University Ave, Palo Alto, CA}
Thrift is a software library and set of code-generation tools developed at
Facebook to expedite development and implementation of efficient and scalable
backend services. Its primary goal is to enable efficient and reliable
communication across programming languages by abstracting the portions of each
language that tend to require the most customization into a common library
that is implemented in each language. Specifically, Thrift allows developers to
define datatypes and service interfaces in a single language-neutral file
and generate all the necessary code to build RPC clients and servers.
This paper details the motivations and design choices we made in Thrift, as
well as some of the more interesting implementation details. It is not
intended to be taken as research, but rather it is an exposition on what we did
and why.
% \category{D.3.3}{Programming Languages}{Language constructs and features}
%Languages, serialization, remote procedure call
%Data description language, interface definition language, remote procedure call
As Facebook's traffic and network structure have scaled, the resource
demands of many operations on the site (i.e. search,
ad selection and delivery, event logging) have presented technical requirements
drastically outside the scope of the LAMP framework. In our implementation of
these services, various programming languages have been selected to
optimize for the right combination of performance, ease and speed of
development, availability of existing libraries, etc. By and large,
Facebook's engineering culture has tended towards choosing the best
tools and implementations available over standardizing on any one
programming language and begrudgingly accepting its inherent limitations.
Given this design choice, we were presented with the challenge of building
a transparent, high-performance bridge across many programming languages.
We found that most available solutions were either too limited, did not offer
sufficient datatype freedom, or suffered from subpar performance.
\footnote{See Appendix A for a discussion of alternative systems.}
The solution that we have implemented combines a language-neutral software
stack implemented across numerous programming languages and an associated code
generation engine that transforms a simple interface and data definition
language into client and server remote procedure call libraries.
Choosing static code generation over a dynamic system allows us to create
validated code that can be run without the need for
any advanced introspective run-time type checking. It is also designed to
be as simple as possible for the developer, who can typically define all
the necessary data structures and interfaces for a complex service in a single
short file.
Surprised that a robust open solution to these relatively common problems
did not yet exist, we committed early on to making the Thrift implementation
open source.
In evaluating the challenges of cross-language interaction in a networked
environment, some key components were identified:
\textit{Types.} A common type system must exist across programming languages
without requiring that the application developer use custom Thrift datatypes
or write their own serialization code. That is,
a C++ programmer should be able to transparently exchange a strongly typed
STL map for a dynamic Python dictionary. Neither
programmer should be forced to write any code below the application layer
to achieve this. Section 2 details the Thrift type system.
\textit{Transport.} Each language must have a common interface to
bidirectional raw data transport. The specifics of how a given
transport is implemented should not matter to the service developer.
The same application code should be able to run against TCP stream sockets,
raw data in memory, or files on disk. Section 3 details the Thrift Transport
\textit{Protocol.} Datatypes must have some way of using the Transport
layer to encode and decode themselves. Again, the application
developer need not be concerned by this layer. Whether the service uses
an XML or binary protocol is immaterial to the application code.
All that matters is that the data can be read and written in a consistent,
deterministic matter. Section 4 details the Thrift Protocol layer.
\textit{Versioning.} For robust services, the involved datatypes must
provide a mechanism for versioning themselves. Specifically,
it should be possible to add or remove fields in an object or alter the
argument list of a function without any interruption in service (or,
worse yet, nasty segmentation faults). Section 5 details Thrift's versioning
\textit{Processors.} Finally, we generate code capable of processing data
streams to accomplish remote procedure calls. Section 6 details the generated
code and TProcessor paradigm.
Section 7 discusses implementation details, and Section 8 describes
our conclusions.
The goal of the Thrift type system is to enable programmers to develop using
completely natively defined types, no matter what programming language they
use. By design, the Thrift type system does not introduce any special dynamic
types or wrapper objects. It also does not require that the developer write
any code for object serialization or transport. The Thrift IDL (Interface
Definition Language) file is
logically a way for developers to annotate their data structures with the
minimal amount of extra information necessary to tell a code generator
how to safely transport the objects across languages.
\subsection{Base Types}
The type system rests upon a few base types. In considering which types to
support, we aimed for clarity and simplicity over abundance, focusing
on the key types available in all programming languages, omitting any
niche types available only in specific languages.
The base types supported by Thrift are:
\item \texttt{bool} A boolean value, true or false
\item \texttt{byte} A signed byte
\item \texttt{i16} A 16-bit signed integer
\item \texttt{i32} A 32-bit signed integer
\item \texttt{i64} A 64-bit signed integer
\item \texttt{double} A 64-bit floating point number
\item \texttt{string} An encoding-agnostic text or binary string
\item \texttt{binary} A byte array representation for blobs
Of particular note is the absence of unsigned integer types. Because these
types have no direct translation to native primitive types in many languages,
the advantages they afford are lost. Further, there is no way to prevent the
application developer in a language like Python from assigning a negative value
to an integer variable, leading to unpredictable behavior. From a design
standpoint, we observed that unsigned integers were very rarely, if ever, used
for arithmetic purposes, but in practice were much more often used as keys or
identifiers. In this case, the sign is irrelevant. Signed integers serve this
same purpose and can be safely cast to their unsigned counterparts (most
commonly in C++) when absolutely necessary.
A Thrift struct defines a common object to be used across languages. A struct
is essentially equivalent to a class in object oriented programming
languages. A struct has a set of strongly typed fields, each with a unique
name identifier. The basic syntax for defining a Thrift struct looks very
similar to a C struct definition. Fields may be annotated with an integer field
identifier (unique to the scope of that struct) and optional default values.
Field identifiers will be automatically assigned if omitted, though they are
strongly encouraged for versioning reasons discussed later.
Thrift containers are strongly typed containers that map to the most commonly
used containers in common programming languages. They are annotated using
the C++ template (or Java Generics) style. There are three types available:
\item \texttt{list<type>} An ordered list of elements. Translates directly into
an STL \texttt{vector}, Java \texttt{ArrayList}, or native array in scripting languages. May
contain duplicates.
\item \texttt{set<type>} An unordered set of unique elements. Translates into
an STL \texttt{set}, Java \texttt{HashSet}, \texttt{set} in Python, or native
dictionary in PHP/Ruby.
\item \texttt{map<type1,type2>} A map of strictly unique keys to values
Translates into an STL \texttt{map}, Java \texttt{HashMap}, PHP associative
array, or Python/Ruby dictionary.
While defaults are provided, the type mappings are not explicitly fixed. Custom
code generator directives have been added to substitute custom types in
destination languages (i.e.
\texttt{hash\_map} or Google's sparse hash map can be used in C++). The
only requirement is that the custom types support all the necessary iteration
primitives. Container elements may be of any valid Thrift type, including other
containers or structs.
struct Example {
1:i32 number=10,
2:i64 bigNumber,
3:double decimals,
4:string name="thrifty"
In the target language, each definition generates a type with two methods,
\texttt{read} and \texttt{write}, which perform serialization and transport
of the objects using a Thrift TProtocol object.
Exceptions are syntactically and functionally equivalent to structs except
that they are declared using the \texttt{exception} keyword instead of the
\texttt{struct} keyword.
The generated objects inherit from an exception base class as appropriate
in each target programming language, in order to seamlessly
integrate with native exception handling in any given
language. Again, the design emphasis is on making the code familiar to the
application developer.
Services are defined using Thrift types. Definition of a service is
semantically equivalent to defining an interface (or a pure virtual abstract
class) in object oriented
programming. The Thrift compiler generates fully functional client and
server stubs that implement the interface. Services are defined as follows:
service <name> {
<returntype> <name>(<arguments>)
[throws (<exceptions>)]
An example:
service StringCache {
void set(1:i32 key, 2:string value),
string get(1:i32 key) throws (1:KeyNotFound knf),
void delete(1:i32 key)
Note that \texttt{void} is a valid type for a function return, in addition to
all other defined Thrift types. Additionally, an \texttt{async} modifier
keyword may be added to a \texttt{void} function, which will generate code that does
not wait for a response from the server. Note that a pure \texttt{void}
function will return a response to the client which guarantees that the
operation has completed on the server side. With \texttt{async} method calls
the client will only be guaranteed that the request succeeded at the
transport layer. (In many transport scenarios this is inherently unreliable
due to the Byzantine Generals' Problem. Therefore, application developers
should take care only to use the async optimization in cases where dropped
method calls are acceptable or the transport is known to be reliable.)
Also of note is the fact that argument lists and exception lists for functions
are implemented as Thrift structs. All three constructs are identical in both
notation and behavior.
The transport layer is used by the generated code to facilitate data transfer.
A key design choice in the implementation of Thrift was to decouple the
transport layer from the code generation layer. Though Thrift is typically
used on top of the TCP/IP stack with streaming sockets as the base layer of
communication, there was no compelling reason to build that constraint into
the system. The performance tradeoff incurred by an abstracted I/O layer
(roughly one virtual method lookup / function call per operation) was
immaterial compared to the cost of actual I/O operations (typically invoking
system calls).
Fundamentally, generated Thrift code only needs to know how to read and
write data. The origin and destination of the data are irrelevant; it may be a
socket, a segment of shared memory, or a file on the local disk. The Thrift
transport interface supports the following methods:
\item \texttt{open} Opens the transport
\item \texttt{close} Closes the transport
\item \texttt{isOpen} Indicates whether the transport is open
\item \texttt{read} Reads from the transport
\item \texttt{write} Writes to the transport
\item \texttt{flush} Forces any pending writes
There are a few additional methods not documented here which are used to aid
in batching reads and optionally signaling the completion of a read or
write operation from the generated code.
In addition to the above
\texttt{TTransport} interface, there is a\\
\texttt{TServerTransport} interface
used to accept or create primitive transport objects. Its interface is as
\item \texttt{open} Opens the transport
\item \texttt{listen} Begins listening for connections
\item \texttt{accept} Returns a new client transport
\item \texttt{close} Closes the transport
The transport interface is designed for simple implementation in any
programming language. New transport mechanisms can be easily defined as needed
by application developers.
The \texttt{TSocket} class is implemented across all target languages. It
provides a common, simple interface to a TCP/IP stream socket.
The \texttt{TFileTransport} is an abstraction of an on-disk file to a data
stream. It can be used to write out a set of incoming Thrift requests to a file
on disk. The on-disk data can then be replayed from the log, either for
post-processing or for reproduction and/or simulation of past events.
The Transport interface is designed to support easy extension using common
OOP techniques, such as composition. Some simple utilities include the
\texttt{TBufferedTransport}, which buffers the writes and reads on an
underlying transport, the \texttt{TFramedTransport}, which transmits data with frame
size headers for chunking optimization or nonblocking operation, and the
\texttt{TMemoryBuffer}, which allows reading and writing directly from the heap
or stack memory owned by the process.
A second major abstraction in Thrift is the separation of data structure from
transport representation. Thrift enforces a certain messaging structure when
transporting data, but it is agnostic to the protocol encoding in use. That is,
it does not matter whether data is encoded as XML, human-readable ASCII, or a
dense binary format as long as the data supports a fixed set of operations
that allow it to be deterministically read and written by generated code.
The Thrift Protocol interface is very straightforward. It fundamentally
supports two things: 1) bidirectional sequenced messaging, and
2) encoding of base types, containers, and structs.
writeMessageBegin(name, type, seq)
writeFieldBegin(name, type, id)
writeMapBegin(ktype, vtype, size)
writeListBegin(etype, size)
writeSetBegin(etype, size)
name, type, seq = readMessageBegin()
name = readStructBegin()
name, type, id = readFieldBegin()
k, v, size = readMapBegin()
etype, size = readListBegin()
etype, size = readSetBegin()
bool = readBool()
byte = readByte()
i16 = readI16()
i32 = readI32()
i64 = readI64()
double = readDouble()
string = readString()
Note that every \texttt{write} function has exactly one \texttt{read} counterpart, with
the exception of \texttt{writeFieldStop()}. This is a special method
that signals the end of a struct. The procedure for reading a struct is to
\texttt{readFieldBegin()} until the stop field is encountered, and then to
\texttt{readStructEnd()}. The
generated code relies upon this call sequence to ensure that everything written by
a protocol encoder can be read by a matching protocol decoder. Further note
that this set of functions is by design more robust than necessary.
For example, \texttt{writeStructEnd()} is not strictly necessary, as the end of
a struct may be implied by the stop field. This method is a convenience for
verbose protocols in which it is cleaner to separate these calls (e.g. a closing
\texttt{</struct>} tag in XML).
Thrift structures are designed to support encoding into a streaming
protocol. The implementation should never need to frame or compute the
entire data length of a structure prior to encoding it. This is critical to
performance in many scenarios. Consider a long list of relatively large
strings. If the protocol interface required reading or writing a list to be an
atomic operation, then the implementation would need to perform a linear pass over the
entire list before encoding any data. However, if the list can be written
as iteration is performed, the corresponding read may begin in parallel,
theoretically offering an end-to-end speedup of $(kN - C)$, where $N$ is the size
of the list, $k$ the cost factor associated with serializing a single
element, and $C$ is fixed offset for the delay between data being written
and becoming available to read.
Similarly, structs do not encode their data lengths a priori. Instead, they are
encoded as a sequence of fields, with each field having a type specifier and a
unique field identifier. Note that the inclusion of type specifiers allows
the protocol to be safely parsed and decoded without any generated code
or access to the original IDL file. Structs are terminated by a field header
with a special \texttt{STOP} type. Because all the basic types can be read
deterministically, all structs (even those containing other structs) can be
read deterministically. The Thrift protocol is self-delimiting without any
framing and regardless of the encoding format.
In situations where streaming is unnecessary or framing is advantageous, it
can be very simply added into the transport layer, using the
\texttt{TFramedTransport} abstraction.
Facebook has implemented and deployed a space-efficient binary protocol which
is used by most backend services. Essentially, it writes all data
in a flat binary format. Integer types are converted to network byte order,
strings are prepended with their byte length, and all message and field headers
are written using the primitive integer serialization constructs. String names
for fields are omitted - when using generated code, field identifiers are
We decided against some extreme storage optimizations (i.e. packing
small integers into ASCII or using a 7-bit continuation format) for the sake
of simplicity and clarity in the code. These alterations can easily be made
if and when we encounter a performance-critical use case that demands them.
Thrift is robust in the face of versioning and data definition changes. This
is critical to enable staged rollouts of changes to deployed services. The
system must be able to support reading of old data from log files, as well as
requests from out-of-date clients to new servers, and vice versa.
\subsection{Field Identifiers}
Versioning in Thrift is implemented via field identifiers. The field header
for every member of a struct in Thrift is encoded with a unique field
identifier. The combination of this field identifier and its type specifier
is used to uniquely identify the field. The Thrift definition language
supports automatic assignment of field identifiers, but it is good
programming practice to always explicitly specify field identifiers.
Identifiers are specified as follows:
struct Example {
1:i32 number=10,
2:i64 bigNumber,
3:double decimals,
4:string name="thrifty"
To avoid conflicts between manually and automatically assigned identifiers,
fields with identifiers omitted are assigned identifiers
decrementing from -1, and the language only supports the manual assignment of
positive identifiers.
When data is being deserialized, the generated code can use these identifiers
to properly identify the field and determine whether it aligns with a field in
its definition file. If a field identifier is not recognized, the generated
code can use the type specifier to skip the unknown field without any error.
Again, this is possible due to the fact that all datatypes are self
Field identifiers can (and should) also be specified in function argument
lists. In fact, argument lists are not only represented as structs on the
backend, but actually share the same code in the compiler frontend. This
allows for version-safe modification of method parameters
service StringCache {
void set(1:i32 key, 2:string value),
string get(1:i32 key) throws (1:KeyNotFound knf),
void delete(1:i32 key)
The syntax for specifying field identifiers was chosen to echo their structure.
Structs can be thought of as a dictionary where the identifiers are keys, and
the values are strongly-typed named fields.
Field identifiers internally use the \texttt{i16} Thrift type. Note, however,
that the \texttt{TProtocol} abstraction may encode identifiers in any format.
When an unexpected field is encountered, it can be safely ignored and
discarded. When an expected field is not found, there must be some way to
signal to the developer that it was not present. This is implemented via an
inner \texttt{isset} structure inside the defined objects. (Isset functionality
is implicit with a \texttt{null} value in PHP, \texttt{None} in Python
and \texttt{nil} in Ruby.) Essentially,
the inner \texttt{isset} object of each Thrift struct contains a boolean value
for each field which denotes whether or not that field is present in the
struct. When a reader receives a struct, it should check for a field being set
before operating directly on it.
class Example {
Example() :
name("thrifty") {}
int32_t number;
int64_t bigNumber;
double decimals;
std::string name;
struct __isset {
__isset() :
name(false) {}
bool number;
bool bigNumber;
bool decimals;
bool name;
} __isset;
\subsection{Case Analysis}
There are four cases in which version mismatches may occur.
\item \textit{Added field, old client, new server.} In this case, the old
client does not send the new field. The new server recognizes that the field
is not set, and implements default behavior for out-of-date requests.
\item \textit{Removed field, old client, new server.} In this case, the old
client sends the removed field. The new server simply ignores it.
\item \textit{Added field, new client, old server.} The new client sends a
field that the old server does not recognize. The old server simply ignores
it and processes as normal.
\item \textit{Removed field, new client, old server.} This is the most
dangerous case, as the old server is unlikely to have suitable default
behavior implemented for the missing field. It is recommended that in this
situation the new server be rolled out prior to the new clients.
\subsection{Protocol/Transport Versioning}
The \texttt{TProtocol} abstractions are also designed to give protocol
implementations the freedom to version themselves in whatever manner they
see fit. Specifically, any protocol implementation is free to send whatever
it likes in the \texttt{writeMessageBegin()} call. It is entirely up to the
implementor how to handle versioning at the protocol level. The key point is
that protocol encoding changes are safely isolated from interface definition
version changes.
Note that the exact same is true of the \texttt{TTransport} interface. For
example, if we wished to add some new checksumming or error detection to the
\texttt{TFileTransport}, we could simply add a version header into the
data it writes to the file in such a way that it would still accept old
log files without the given header.
\section{RPC Implementation}
The last core interface in the Thrift design is the \texttt{TProcessor},
perhaps the most simple of the constructs. The interface is as follows:
interface TProcessor {
bool process(TProtocol in, TProtocol out)
throws TException
The key design idea here is that the complex systems we build can fundamentally
be broken down into agents or services that operate on inputs and outputs. In
most cases, there is actually just one input and output (an RPC client) that
needs handling.
\subsection{Generated Code}
When a service is defined, we generate a
\texttt{TProcessor} instance capable of handling RPC requests to that service,
using a few helpers. The fundamental structure (illustrated in pseudo-C++) is
as follows:
=> Service.cpp
interface ServiceIf
class ServiceClient : virtual ServiceIf
TProtocol in
TProtocol out
class ServiceProcessor : TProcessor
ServiceIf handler
class ServiceHandler : virtual ServiceIf
TServer(TProcessor processor,
TServerTransport transport,
TTransportFactory tfactory,
TProtocolFactory pfactory)
From the Thrift definition file, we generate the virtual service interface.
A client class is generated, which implements the interface and
uses two \texttt{TProtocol} instances to perform the I/O operations. The
generated processor implements the \texttt{TProcessor} interface. The generated
code has all the logic to handle RPC invocations via the \texttt{process()}
call, and takes as a parameter an instance of the service interface, as
implemented by the application developer.
The user provides an implementation of the application interface in separate,
non-generated source code.
Finally, the Thrift core libraries provide a \texttt{TServer} abstraction.
The \texttt{TServer} object generally works as follows.
\item Use the \texttt{TServerTransport} to get a \texttt{TTransport}
\item Use the \texttt{TTransportFactory} to optionally convert the primitive
transport into a suitable application transport (typically the
\texttt{TBufferedTransportFactory} is used here)
\item Use the \texttt{TProtocolFactory} to create an input and output protocol
for the \texttt{TTransport}
\item Invoke the \texttt{process()} method of the \texttt{TProcessor} object
The layers are appropriately separated such that the server code needs to know
nothing about any of the transports, encodings, or applications in play. The
server encapsulates the logic around connection handling, threading, etc.
while the processor deals with RPC. The only code written by the application
developer lives in the definitional Thrift file and the interface
Facebook has deployed multiple \texttt{TServer} implementations, including
the single-threaded \texttt{TSimpleServer}, thread-per-connection
\texttt{TThreadedServer}, and thread-pooling \texttt{TThreadPoolServer}.
The \texttt{TProcessor} interface is very general by design. There is no
requirement that a \texttt{TServer} take a generated \texttt{TProcessor}
object. Thrift allows the application developer to easily write any type of
server that operates on \texttt{TProtocol} objects (for instance, a server
could simply stream a certain type of object without any actual RPC method
\section{Implementation Details}
\subsection{Target Languages}
Thrift currently supports five target languages: C++, Java, Python, Ruby, and
PHP. At Facebook, we have deployed servers predominantly in C++, Java, and
Python. Thrift services implemented in PHP have also been embedded into the
Apache web server, providing transparent backend access to many of our
frontend constructs using a \texttt{THttpClient} implementation of the
\texttt{TTransport} interface.
Though Thrift was explicitly designed to be much more efficient and robust
than typical web technologies, as we were designing our XML-based REST web
services API we noticed that Thrift could be easily used to define our
service interface. Though we do not currently employ SOAP envelopes (in the
authors' opinions there is already far too much repetitive enterprise Java
software to do that sort of thing), we were able to quickly extend Thrift to
generate XML Schema Definition files for our service, as well as a framework
for versioning different implementations of our web service. Though public
web services are admittedly tangential to Thrift's core use case and design,
Thrift facilitated rapid iteration and affords us the ability to quickly
migrate our entire XML-based web service onto a higher performance system
should the need arise.
\subsection{Generated Structs}
We made a conscious decision to make our generated structs as transparent as
possible. All fields are publicly accessible; there are no \texttt{set()} and
\texttt{get()} methods. Similarly, use of the \texttt{isset} object is not
enforced. We do not include any \texttt{FieldNotSetException} construct.
Developers have the option to use these fields to write more robust code, but
the system is robust to the developer ignoring the \texttt{isset} construct
entirely and will provide suitable default behavior in all cases.
This choice was motivated by the desire to ease application development. Our stated
goal is not to make developers learn a rich new library in their language of
choice, but rather to generate code that allow them to work with the constructs
that are most familiar in each language.
We also made the \texttt{read()} and \texttt{write()} methods of the generated
objects public so that the objects can be used outside of the context
of RPC clients and servers. Thrift is a useful tool simply for generating
objects that are easily serializable across programming languages.
\subsection{RPC Method Identification}
Method calls in RPC are implemented by sending the method name as a string. One
issue with this approach is that longer method names require more bandwidth.
We experimented with using fixed-size hashes to identify methods, but in the
end concluded that the savings were not worth the headaches incurred. Reliably
dealing with conflicts across versions of an interface definition file is
impossible without a meta-storage system (i.e. to generate non-conflicting
hashes for the current version of a file, we would have to know about all
conflicts that ever existed in any previous version of the file).
We wanted to avoid too many unnecessary string comparisons upon
method invocation. To deal with this, we generate maps from strings to function
pointers, so that invocation is effectively accomplished via a constant-time
hash lookup in the common case. This requires the use of a couple interesting
code constructs. Because Java does not have function pointers, process
functions are all private member classes implementing a common interface.
private class ping implements ProcessFunction {
public void process(int seqid,
TProtocol iprot,
TProtocol oprot)
throws TException
{ ...}
HashMap<String,ProcessFunction> processMap_ =
new HashMap<String,ProcessFunction>();
In C++, we use a relatively esoteric language construct: member function
void (ExampleServiceProcessor::*)(int32_t,
Using these techniques, the cost of string processing is minimized, and we
reap the benefit of being able to easily debug corrupt or misunderstood data by
inspecting it for known string method names.
\subsection{Servers and Multithreading}
Thrift services require basic multithreading to handle simultaneous
requests from multiple clients. For the Python and Java implementations of
Thrift server logic, the standard threading libraries distributed with the
languages provide adequate support. For the C++ implementation, no standard multithread runtime
library exists. Specifically, robust, lightweight, and portable
thread manager and timer class implementations do not exist. We investigated
existing implementations, namely \texttt{boost::thread},
\texttt{boost::threadpool}, \texttt{ACE\_Thread\_Manager} and
While \texttt{boost::threads}\cite{boost.threads} provides clean,
lightweight and robust implementations of multi-thread primitives (mutexes,
conditions, threads) it does not provide a thread manager or timer
\texttt{boost::threadpool}\cite{boost.threadpool} also looked promising but
was not far enough along for our purposes. We wanted to limit the dependency on
third-party libraries as much as possible. Because\\
\texttt{boost::threadpool} is
not a pure template library and requires runtime libraries and because it is
not yet part of the official Boost distribution we felt it was not ready for
use in Thrift. As \texttt{boost::threadpool} evolves and especially if it is
added to the Boost distribution we may reconsider our decision to not use it.
ACE has both a thread manager and timer class in addition to multi-thread
primitives. The biggest problem with ACE is that it is ACE. Unlike Boost, ACE
API quality is poor. Everything in ACE has large numbers of dependencies on
everything else in ACE - thus forcing developers to throw out standard
classes, such as STL collections, in favor of ACE's homebrewed implementations. In
addition, unlike Boost, ACE implementations demonstrate little understanding
of the power and pitfalls of C++ programming and take no advantage of modern
templating techniques to ensure compile time safety and reasonable compiler
error messages. For all these reasons, ACE was rejected. Instead, we chose
to implement our own library, described in the following sections.
\subsection{Thread Primitives}
The Thrift thread libraries are implemented in the namespace\\
\texttt{facebook::thrift::concurrency} and have three components:
\item primitives
\item thread pool manager
\item timer manager
As mentioned above, we were hesitant to introduce any additional dependencies
on Thrift. We decided to use \texttt{boost::shared\_ptr} because it is so
useful for multithreaded application, it requires no link-time or
runtime libraries (i.e. it is a pure template library) and it is due
to become part of the C++0x standard.
We implement standard \texttt{Mutex} and \texttt{Condition} classes, and a
\texttt{Monitor} class. The latter is simply a combination of a mutex and
condition variable and is analogous to the \texttt{Monitor} implementation provided for
the Java \texttt{Object} class. This is also sometimes referred to as a barrier. We
provide a \texttt{Synchronized} guard class to allow Java-like synchronized blocks.
This is just a bit of syntactic sugar, but, like its Java counterpart, clearly
delimits critical sections of code. Unlike its Java counterpart, we still
have the ability to programmatically lock, unlock, block, and signal monitors.
void run() {
{Synchronized s(manager->monitor);
if (manager->state == TimerManager::STARTING) {
manager->state = TimerManager::STARTED;
We again borrowed from Java the distinction between a thread and a runnable
class. A \texttt{Thread} is the actual schedulable object. The
\texttt{Runnable} is the logic to execute within the thread.
The \texttt{Thread} implementation deals with all the platform-specific thread
creation and destruction issues, while the \texttt{Runnable} implementation deals
with the application-specific per-thread logic. The benefit of this approach
is that developers can easily subclass the Runnable class without pulling in
platform-specific super-classes.
\subsection{Thread, Runnable, and shared\_ptr}
We use \texttt{boost::shared\_ptr} throughout the \texttt{ThreadManager} and
\texttt{TimerManager} implementations to guarantee cleanup of dead objects that can
be accessed by multiple threads. For \texttt{Thread} class implementations,
\texttt{boost::shared\_ptr} usage requires particular attention to make sure
\texttt{Thread} objects are neither leaked nor dereferenced prematurely while
creating and shutting down threads.
Thread creation requires calling into a C library. (In our case the POSIX
thread library, \texttt{libpthread}, but the same would be true for WIN32 threads).
Typically, the OS makes few, if any, guarantees about when \texttt{ThreadMain}, a C thread's entry-point function, will be called. Therefore, it is
possible that our thread create call,
\texttt{ThreadFactory::newThread()} could return to the caller
well before that time. To ensure that the returned \texttt{Thread} object is not
prematurely cleaned up if the caller gives up its reference prior to the
\texttt{ThreadMain} call, the \texttt{Thread} object makes a weak reference to
itself in its \texttt{start} method.
With the weak reference in hand the \texttt{ThreadMain} function can attempt to get
a strong reference before entering the \texttt{Runnable::run} method of the
\texttt{Runnable} object bound to the \texttt{Thread}. If no strong references to the
thread are obtained between exiting \texttt{Thread::start} and entering \texttt{ThreadMain}, the weak reference returns \texttt{null} and the function
exits immediately.
The need for the \texttt{Thread} to make a weak reference to itself has a
significant impact on the API. Since references are managed through the
\texttt{boost::shared\_ptr} templates, the \texttt{Thread} object must have a reference
to itself wrapped by the same \texttt{boost::shared\_ptr} envelope that is returned
to the caller. This necessitated the use of the factory pattern.
\texttt{ThreadFactory} creates the raw \texttt{Thread} object and a
\texttt{boost::shared\_ptr} wrapper, and calls a private helper method of the class
implementing the \texttt{Thread} interface (in this case, \texttt{PosixThread::weakRef})
to allow it to make add weak reference to itself through the
\texttt{boost::shared\_ptr} envelope.
\texttt{Thread} and \texttt{Runnable} objects reference each other. A \texttt{Runnable}
object may need to know about the thread in which it is executing, and a Thread, obviously,
needs to know what \texttt{Runnable} object it is hosting. This interdependency is
further complicated because the lifecycle of each object is independent of the
other. An application may create a set of \texttt{Runnable} object to be reused in different threads, or it may create and forget a \texttt{Runnable} object
once a thread has been created and started for it.
The \texttt{Thread} class takes a \texttt{boost::shared\_ptr} reference to the hosted
\texttt{Runnable} object in its constructor, while the \texttt{Runnable} class has an
explicit \texttt{thread} method to allow explicit binding of the hosted thread.
\texttt{ThreadFactory::newThread} binds the objects to each other.
\texttt{ThreadManager} creates a pool of worker threads and
allows applications to schedule tasks for execution as free worker threads
become available. The \texttt{ThreadManager} does not implement dynamic
thread pool resizing, but provides primitives so that applications can add
and remove threads based on load. This approach was chosen because
implementing load metrics and thread pool size is very application
specific. For example some applications may want to adjust pool size based
on running-average of work arrival rates that are measured via polled
samples. Others may simply wish to react immediately to work-queue
depth high and low water marks. Rather than trying to create a complex
API abstract enough to capture these different approaches, we
simply leave it up to the particular application and provide the
primitives to enact the desired policy and sample current status.
\texttt{TimerManager} allows applications to schedule
\texttt{Runnable} objects for execution at some point in the future. Its specific task
is to allows applications to sample \texttt{ThreadManager} load at regular
intervals and make changes to the thread pool size based on application policy.
Of course, it can be used to generate any number of timer or alarm events.
The default implementation of \texttt{TimerManager} uses a single thread to
execute expired \texttt{Runnable} objects. Thus, if a timer operation needs to
do a large amount of work and especially if it needs to do blocking I/O,
that should be done in a separate thread.
\subsection{Nonblocking Operation}
Though the Thrift transport interfaces map more directly to a blocking I/O
model, we have implemented a high performance \texttt{TNonBlockingServer}
in C++ based on \texttt{libevent} and the \texttt{TFramedTransport}. We
implemented this by moving all I/O into one tight event loop using a
state machine. Essentially, the event loop reads framed requests into
\texttt{TMemoryBuffer} objects. Once entire requests are ready, they are
dispatched to the \texttt{TProcessor} object which can read directly from
the data in memory.
The Thrift compiler is implemented in C++ using standard \texttt{lex}/\texttt{yacc}
lexing and parsing. Though it could have been implemented with fewer
lines of code in another language (i.e. Python Lex-Yacc (PLY) or \texttt{ocamlyacc}), using C++
forces explicit definition of the language constructs. Strongly typing the
parse tree elements (debatably) makes the code more approachable for new
Code generation is done using two passes. The first pass looks only for
include files and type definitions. Type definitions are not checked during
this phase, since they may depend upon include files. All included files
are sequentially scanned in a first pass. Once the include tree has been
resolved, a second pass over all files is taken that inserts type definitions
into the parse tree and raises an error on any undefined types. The program is
then generated against the parse tree.
Due to inherent complexities and potential for circular dependencies,
we explicitly disallow forward declaration. Two Thrift structs cannot
each contain an instance of the other. (Since we do not allow \texttt{null}
struct instances in the generated C++ code, this would actually be impossible.)
The \texttt{TFileTransport} logs Thrift requests/structs by
framing incoming data with its length and writing it out to disk.
Using a framed on-disk format allows for better error checking and
helps with the processing of a finite number of discrete events. The\\
\texttt{TFileWriterTransport} uses a system of swapping in-memory buffers
to ensure good performance while logging large amounts of data.
A Thrift log file is split up into chunks of a specified size; logged messages
are not allowed to cross chunk boundaries. A message that would cross a chunk
boundary will cause padding to be added until the end of the chunk and the
first byte of the message are aligned to the beginning of the next chunk.
Partitioning the file into chunks makes it possible to read and interpret data
from a particular point in the file.
\section{Facebook Thrift Services}
Thrift has been employed in a large number of applications at Facebook, including
search, logging, mobile, ads and the developer platform. Two specific usages are discussed below.
Thrift is used as the underlying protocol and transport layer for the Facebook Search service.
The multi-language code generation is well suited for search because it allows for application
development in an efficient server side language (C++) and allows the Facebook PHP-based web application
to make calls to the search service using Thrift PHP libraries. There is also a large
variety of search stats, deployment and testing functionality that is built on top
of generated Python code. Additionally, the Thrift log file format is
used as a redo log for providing real-time search index updates. Thrift has allowed the
search team to leverage each language for its strengths and to develop code at a rapid pace.
The Thrift \texttt{TFileTransport} functionality is used for structured logging. Each
service function definition along with its parameters can be considered to be
a structured log entry identified by the function name. This log can then be used for
a variety of purposes, including inline and offline processing, stats aggregation and as a redo log.
Thrift has enabled Facebook to build scalable backend
services efficiently by enabling engineers to divide and conquer. Application
developers can focus on application code without worrying about the
sockets layer. We avoid duplicated work by writing buffering and I/O logic
in one place, rather than interspersing it in each application.
Thrift has been employed in a wide variety of applications at Facebook,
including search, logging, mobile, ads, and the developer platform. We have
found that the marginal performance cost incurred by an extra layer of
software abstraction is far eclipsed by the gains in developer efficiency and
systems reliability.
\section{Similar Systems}
The following are software systems similar to Thrift. Each is (very!) briefly
\item \textit{SOAP.} XML-based. Designed for web services via HTTP, excessive
XML parsing overhead.
\item \textit{CORBA.} Relatively comprehensive, debatably overdesigned and
heavyweight. Comparably cumbersome software installation.
\item \textit{COM.} Embraced mainly in Windows client software. Not an entirely
open solution.
\item \textit{Pillar.} Lightweight and high-performance, but missing versioning
and abstraction.
\item \textit{Protocol Buffers.} Closed-source, owned by Google. Described in
Sawzall paper.
Many thanks for feedback on Thrift (and extreme trial by fire) are due to
Martin Smith, Karl Voskuil and Yishan Wong.
Thrift is a successor to Pillar, a similar system developed
by Adam D'Angelo, first while at Caltech and continued later at Facebook.
Thrift simply would not have happened without Adam's insights.
Kempf, William,
Henkel, Philipp,