Riser
RISER is a library of Ruby Infrastructure for cooperative multi-thread/multi-process SERver.
This library is useful for the following.
- To make a server with tcp/ip or unix domain socket
- To select a method to execute server from:
- Single process multi-thread
- Preforked multi-process multi-thread
- To make a daemon that will be controlled by signal(2)s
- To separate the object not divided into multiple processes from server process(es) into backend service process
This library supposes that the user is familiar with the unix process model and socket programming.
Installation
Add this line to your application's Gemfile:
gem 'riser'
And then execute:
$ bundle
Or install it yourself as:
$ gem install riser
Usage
Simple Server Example
An example of a simple server is as follows.
require 'riser'
require 'socket'
server = Riser::SocketServer.new
server.dispatch{|socket|
while (line = socket.gets)
socket.write(line)
end
}
server_socket = TCPServer.new('localhost', 5000)
server.start(server_socket)
This simple server is an echo server that accepts connections at port
number 5000 on localhost and returns the input line as is. The object
of Riser::SocketServer
is the core of riser. What this example does
is as follows.
- Create a new server object of
Riser::SocketServer
. - Register
dispatch
callback to the server object. - Open a tcp/ip server socket.
- Pass the server socket to the server object and
start
the server.
In this example tcp/ip socket is used, but the server will also work
with unix domain socket. By rewriting the dispatch
callback you can
make the server do what you want. Although this example is
simplified, error handling is actually required. If an exception is
thrown to outside of the dispatch
callback, the server stops, so it
should be avoided. See the
'halo.rb'
example for practical code.
By default, the server performs dispatch
callback on multi-thread.
On Linux, you can see that running the server with 4 threads
({ruby}
) by executing pstree
command.
$ pstree -ap
...
| `-bash,23355
| `-ruby,30674 simple_server.rb
| |-{ruby},30675
| |-{ruby},30676
| |-{ruby},30677
| `-{ruby},30678
...
Server Attributes
The server has attributes, and setting attributes changes the behavior
of the server. Let's take an example process_num
attribute. The
process_num
attribute is 0
by default, but by setting it will run
the server with multi-process.
In the following example, process_num
is set to 2
.
Others are the same as simple server example.
require 'riser'
require 'socket'
server = Riser::SocketServer.new
server.process_num = 2
server.dispatch{|socket|
while (line = socket.gets)
socket.write(line)
end
}
server_socket = TCPServer.new('localhost', 5000)
server.start(server_socket)
Running the example will hardly change the appearance, but the server
is running with multi-process. You can see that the server is running
in multiple processes with pstree
command.
$ pstree -ap
...
| `-bash,23355
| `-ruby,31283 multiproc_server.rb
| |-ruby,31284 multiproc_server.rb
| | |-{ruby},31285
| | |-{ruby},31286
| | |-{ruby},31287
| | `-{ruby},31288
| |-ruby,31289 multiproc_server.rb
| | |-{ruby},31292
| | |-{ruby},31293
| | |-{ruby},31294
| | `-{ruby},31295
| |-{ruby},31290
| `-{ruby},31291
...
There are 2 child processes (|-ruby
) under the parent process
( `-ruby
) having 2 threads , and 4 threads are running for each
child process. The architecture of riser's multi-process server is
the passing file descriptors between parent-child processes. The
parent process accepts the connection and passes it to threads, each
thread passes the connection to the child process, and dispatch
callback is performed in the thread of each child process.
In addition to process_num
, the server object has various attributes
such as thread_num
. See the source code
(server.rb: Riser::SocketServer)
for details of other attributes.
Daemon
Riser provids the function to daemonize server. By daemonizing the server, the server will be able to receive signal(2)s and restart. An example of a simple daemon is as follows.
require 'riser'
Riser::Daemon.start_daemon(daemonize: true,
daemon_name: 'simple_daemon',
status_file: 'simple_daemon.pid',
listen_address: 'localhost:5000'
) {|server|
server.dispatch{|socket|
while (line = socket.gets)
socket.write(line)
end
}
}
To daemonize the server, use the module function of
Riser::Daemon.start_daemon
. The start_daemon
function takes
parameters in a hash table and works. The works of start_daemon
are
as follows.
- Daemonize the server process (
daemonize: true
). - Output syslog(2) identified with
simple_daemon
(daemon_name: 'simple_daemon'
). - Output process id to the file of
simple_daemon.pid
and lock it exclusively (status_file: 'simple_daemon.pid'
). - Open the tcp/ip server socket of
localhost:5000
(listen_address: 'localhost:5000'
). - Create a server object and pass it to the block, and you set up the
server object in the block, then
start
the server object.
A command prompt is displayed as soon as you start the daemon, but the
daemon runs in the background and logs to syslog(2). Daemonization is
the result of daemonaize: true
. If daemonaize: false
is set, the
process is not daemonized, starts in foreground and logs to standard
output. This is useful for debugging daemon.
Looking at the process of the daemon with pstree
command is as
follows.
$ pstree -ap
init,1 ro
|-ruby,32187 simple_daemon.rb
| `-ruby,32188 simple_daemon.rb
| |-{ruby},32189
| |-{ruby},32190
| |-{ruby},32191
| `-{ruby},32192
...
The daemon process is running as the parent of the server process. And the daemon process is running independently under the init(8) process. The daemon process monitors the server process and restarts when the server process dies. Also, the daemon process receives some signal(2)s and stops or restarts the server process, and does other things.
Signal(2)s and Other Daemon Parameters
By default, the daemon is able to receive the following signal(2)s.
signal(2) | daemon's action |
start_daemon parameter |
---|---|---|
TERM |
stop server gracefully | signal_stop_graceful |
INT |
stop server forcedly | signal_stop_forced |
HUP |
restart server gracefully | signal_restart_graceful |
QUIT |
restart server forcedly | signal_restart_forced |
USR1 |
get queue stat and reset queue stat | signal_stat_get_and_reset |
USR2 |
get queue stat and no reset queue stat | signal_stat_get_no_reset |
WINCH |
stop queue stat | signal_stat_stop |
By setting the parameters of the start_daemon
, you can change the
signal(2) which triggers the action. Setting the parameter to nil
will disable the action. 'Queue stat' is explained later.
The start_daemon
has other parameters. See the source code
(daemon.rb: Riser::Daemon::DEFAULT)
for details of other parameters.
Server Callbacks
The server object is able to register callbacks other than dispatch
.
The list of server objects' callbacks is as follows.
callback | description |
---|---|
before_start{|server_socket| ...} |
performed before starting the server. in a multi-process server, it is performed in the parent process. |
at_fork{ ... } |
performed after fork(2)ing on the multi-process server. it is performed in the child process. |
at_stop{|stop_state| ... } |
performed when a stop signal(2) is received. in a multi-process server, it is performed in the child process. |
at_stat{|stat_info| ... } |
performed when 'get stat' signal(2) is received. |
preprocess{ ... } |
performed before starting 'dispatch loop'. in a multi-process server, it is performed in the child process. |
postprocess{ ... } |
performed after 'dispatch loop' is finished. in a multi-process server, it is performed in the child process. |
after_stop{ ... } |
performed after the server stop. in a multi-process server, it is performed in the parent process. |
dispatch{|socket| ... } |
known dispatch callback. in a multi-process server, it is performed in the child process. |
It seems necessary to explain the at_stat
callback. Riser uses
queues to distribute connections to threads and processes, and it is
possible to get statistics information on queues. With USR1
and
USR2
signal(2)s, you can start collecting queue statistics
information and get it. At that time the at_stat
callback is called
and used to write queue statistics informations to log etc. With the
WINCH
signal(2), you can stop collecting queue statistics
information.
For a example of how to use callbacks, see the source code of the 'halo.rb' example.
Server Utilities
Riser provides some useful utilities to write a server.
utility | description |
---|---|
Riser::ReadPoll |
monitor I/O timeout. |
Riser::WriteBufferStream |
buffer I/O writes. |
Riser::LoggingStream |
log I/O read / write. |
For a example of how to use utilities, see the source code of the 'halo.rb' example. Also utilities are simple, so check the source codes of 'poll.rb' and 'stream.rb'.
TLS Server
With OpenSSL, the riser is able to provide a TLS server. To provide a TLS server you need a certificate and a private key. An example of a simple TLS server is as follows.
require 'openssl'
require 'riser'
cert_path = ARGV.shift or abort('need for server certificate file')
pkey_path = ARGV.shift or abort('need for server private key file')
Riser::Daemon.start_daemon(daemonize: false,
daemon_name: 'simple_tls',
listen_address: 'localhost:5000'
) {|server|
ssl_context = OpenSSL::SSL::SSLContext.new
ssl_context.cert = OpenSSL::X509::Certificate.new(File.read(cert_path))
ssl_context.key = OpenSSL::PKey.read(File.read(pkey_path))
server.dispatch{|socket|
ssl_socket = OpenSSL::SSL::SSLSocket.new(socket, ssl_context)
ssl_socket.accept
while (line = ssl_socket.gets)
ssl_socket.write(line)
end
ssl_socket.close
}
}
An example of the result of connecting to the TLS server from OpenSSL client is as follows.
$ openssl s_client -CAfile local_ca.cert -connect localhost:5000
CONNECTED(00000003)
depth=1 C = JP, ST = Tokyo, L = Tokyo, O = Private, OU = Home, CN = *
verify return:1
depth=0 C = JP, ST = Tokyo, L = Tokyo, O = Private, OU = Home, CN = localhost
verify return:1
---
Certificate chain
0 s:/C=JP/ST=Tokyo/L=Tokyo/O=Private/OU=Home/CN=localhost
i:/C=JP/ST=Tokyo/L=Tokyo/O=Private/OU=Home/CN=*
---
Server certificate
-----BEGIN CERTIFICATE-----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-----END CERTIFICATE-----
subject=/C=JP/ST=Tokyo/L=Tokyo/O=Private/OU=Home/CN=localhost
issuer=/C=JP/ST=Tokyo/L=Tokyo/O=Private/OU=Home/CN=*
---
No client certificate CA names sent
Peer signing digest: SHA512
Server Temp Key: X25519, 253 bits
---
SSL handshake has read 1453 bytes and written 269 bytes
Verification: OK
---
New, TLSv1.2, Cipher is ECDHE-RSA-AES256-GCM-SHA384
Server public key is 2048 bit
Secure Renegotiation IS supported
Compression: NONE
Expansion: NONE
No ALPN negotiated
SSL-Session:
Protocol : TLSv1.2
Cipher : ECDHE-RSA-AES256-GCM-SHA384
Session-ID: 43427207A2C36F807AF5BDCB69EF26F18758A5BAC5C4867C04B17E1C1F6CAE9D
Session-ID-ctx:
Master-Key: 7BE6C8E0108A6A2F9B2B6AC4DB8360EE375A950D2EB4CB2B259125FB17BE74F00120F7E7290B7137E16F665F44D8AD20
PSK identity: None
PSK identity hint: None
SRP username: None
TLS session ticket lifetime hint: 7200 (seconds)
TLS session ticket:
0000 - af bd 28 5e ca d4 ae 61-38 63 ff 68 84 2b 51 13 ..(^...a8c.h.+Q.
0010 - d3 c5 7f c7 72 be f5 5c-bd 9e fb f3 88 61 83 01 ....r..\.....a..
0020 - a5 11 fe 45 14 c1 9c 9b-79 7b 34 87 c1 66 e1 cd ...E....y{4..f..
0030 - 7d f4 ac 62 6e 25 53 c5-35 b2 b2 2c 3c b9 af 89 }..bn%S.5..,<...
0040 - cf 11 1d 9c 42 5a 75 86-d1 6d 49 fc e9 6a 39 f0 ....BZu..mI..j9.
0050 - fb cf 7d 9a 60 52 10 ad-a3 15 1b ba 00 32 67 e8 ..}.`R.......2g.
0060 - 03 ea 74 49 17 46 d8 a2-41 45 17 9d 2c ec 7f 3f ..tI.F..AE..,..?
0070 - 89 eb 7e 4a 05 10 a3 81-d2 16 ce c7 da 7d c6 5a ..~J.........}.Z
0080 - 9c 50 de a5 ce 8e ca 58-af 0b 94 d2 2a c2 56 da .P.....X....*.V.
0090 - 00 05 b9 87 3c 9c 0e 53-70 c2 59 24 ef 0b 0a f3 ....<..Sp.Y$....
Start Time: 1549358604
Timeout : 7200 (sec)
Verify return code: 0 (ok)
Extended master secret: yes
---
foo
foo
bar
bar
DONE
dRuby Services
Riser has a mechanism that runs the object in a separate process from the server process. This mechanism pools the dRuby server processes and distributes the object to them in the following 3 patterns.
pattern | description |
---|---|
any process | run the object with a randomly picked process. |
single process | always run the object in the same process. |
sticky process | run the object in the same process for each specific key. |
A simple example of how this mechanism works is as follows.
require 'riser'
Riser::Daemon.start_daemon(daemonize: false,
daemon_name: 'simple_services',
listen_address: 'localhost:8000'
) {|server|
services = Riser::DRbServices.new(4)
services.add_any_process_service(:pid_any, proc{ $$ })
services.add_single_process_service(:pid_single, proc{ $$ })
services.add_sticky_process_service(:pid_stickty, proc{|key| $$ })
server.process_num = 2
server.before_start{|server_socket|
services.start_server
}
server.at_fork{
services.detach_server
}
server.preprocess{
services.start_client
}
server.dispatch{|socket|
if (line = socket.gets) then
method, uri, _version = line.split
while (line = socket.gets)
line.strip.empty? and break
end
if (method == 'GET') then
socket << "HTTP/1.0 200 OK\r\n"
socket << "Content-Type: text/plain\r\n"
socket << "\r\n"
path, query = uri.split('?', 2)
case (path)
when '/any'
socket << 'pid: ' << services.call_service(:pid_any) << "\n"
when '/single'
socket << 'pid: ' << services.call_service(:pid_single) << "\n"
when '/sticky'
key = query || 'default'
socket << 'key: ' << key << "\n"
socket << 'pid: ' << services.call_service(:pid_stickty, key) << "\n"
else
socket << "unknown path: #{path}\n"
end
end
end
}
server.after_stop{
services.stop_server
}
}
Riser::DRbServices
is the mechanism for distributing objects. What
this example does is as follows.
- Create a object of
Riser::DRbServices
to pool 4 dRuby server processes (Riser::DRbServices.new(4)
). - Add objects to run in dRuby server process
(
add_..._process_service
). In this example it is added that the procedures that returns the process id to see the process to be distributed the object. - Start dRuby server process with
start_server
. In the case of a multi-process server, it is necessary to executestart_server
in the parent process, so it is executed atbefore_start
callback. - In the case of a multi-process server, execute
detach_server
atat_fork
callback to release unnecessary resources in the child process. - Start dRuby client with
start_client
. In the case of a multi-process server, it is necessary to executestart_client
in the child process, so it is executed atpreprocess
callback. - Add the processing of the web service at
dispatch
callback. The procedures added toRiser::DRbServices
are able to be called bycall_service
. For object other than procedure, useget_service
. - Stop dRuby server process with
stop_server
. In the case of a multi-process server, it is necessary to executestop_server
in the parent process, so it is executed atafter_stop
callback.
In this example, you can see the dRuby process distribution with a
simple web service. Looking at the process of this example with
pstree
command is as follows.
$ pstree -ap
...
| `-bash,23355
| `-ruby,3177 simple_services.rb
| `-ruby,3178 simple_services.rb
| |-ruby,3179 simple_services.rb
| | |-{ruby},3180
| | |-{ruby},3189
| | `-{ruby},3198
| |-ruby,3181 simple_services.rb
| | |-{ruby},3182
| | |-{ruby},3194
| | `-{ruby},3202
| |-ruby,3183 simple_services.rb
| | |-{ruby},3184
| | |-{ruby},3197
| | `-{ruby},3207
| |-ruby,3185 simple_services.rb
| | |-{ruby},3186
| | |-{ruby},3201
| | `-{ruby},3211
| |-ruby,3187 simple_services.rb
| | |-{ruby},3188
| | |-{ruby},3191
| | |-{ruby},3195
| | |-{ruby},3199
| | |-{ruby},3203
| | |-{ruby},3205
| | |-{ruby},3206
| | |-{ruby},3208
| | `-{ruby},3209
| |-ruby,3190 simple_services.rb
| | |-{ruby},3196
| | |-{ruby},3200
| | |-{ruby},3204
| | |-{ruby},3210
| | |-{ruby},3212
| | |-{ruby},3213
| | |-{ruby},3214
| | |-{ruby},3215
| | `-{ruby},3216
| |-{ruby},3192
| `-{ruby},3193
...
In addition to the 2 server child processes with many threads, there are 4 child processes. These 4 child processes are dRuby server processes. See the web service's result of 'any process' pattern.
$ curl http://localhost:8000/any
pid: 3181
$ curl http://localhost:8000/any
pid: 3179
$ curl http://localhost:8000/any
pid: 3183
$ curl http://localhost:8000/any
pid: 3181
In the 'any process' pattern, process ids are dispersed. Next, see the web service's result of 'single process' pattern.
$ curl http://localhost:8000/single
pid: 3179
$ curl http://localhost:8000/single
pid: 3179
$ curl http://localhost:8000/single
pid: 3179
$ curl http://localhost:8000/single
pid: 3179
In the 'single process' pattern, process id is always same. Last, see the web service's result of 'sticky process' pattern.
$ curl http://localhost:8000/sticky
key: default
pid: 3181
$ curl http://localhost:8000/sticky
key: default
pid: 3181
$ curl http://localhost:8000/sticky?foo
key: foo
pid: 3179
$ curl http://localhost:8000/sticky?foo
key: foo
pid: 3179
$ curl http://localhost:8000/sticky?bar
key: bar
pid: 3185
$ curl http://localhost:8000/sticky?bar
key: bar
pid: 3185
In the 'sticky process' pattern, the same process id will be given for each key.
Local Services
Since dRuby's remote process call has overhead, riser is able to
transparently switch Riser::DRbServices
to local process call. An
example of a local process call is as follows.
require 'riser'
Riser::Daemon.start_daemon(daemonize: false,
daemon_name: 'local_services',
listen_address: 'localhost:8000'
) {|server|
services = Riser::DRbServices.new(0)
services.add_any_process_service(:pid_any, proc{ $$ })
services.add_single_process_service(:pid_single, proc{ $$ })
services.add_sticky_process_service(:pid_stickty, proc{|key| $$ })
server.process_num = 0
server.before_start{|server_socket|
services.start_server
}
server.at_fork{
services.detach_server
}
server.preprocess{
services.start_client
}
server.dispatch{|socket|
if (line = socket.gets) then
method, uri, _version = line.split
while (line = socket.gets)
line.strip.empty? and break
end
if (method == 'GET') then
socket << "HTTP/1.0 200 OK\r\n"
socket << "Content-Type: text/plain\r\n"
socket << "\r\n"
path, query = uri.split('?', 2)
case (path)
when '/any'
socket << 'pid: ' << services.call_service(:pid_any) << "\n"
when '/single'
socket << 'pid: ' << services.call_service(:pid_single) << "\n"
when '/sticky'
key = query || 'default'
socket << 'key: ' << key << "\n"
socket << 'pid: ' << services.call_service(:pid_stickty, key) << "\n"
else
socket << "unknown path: #{path}\n"
end
end
end
}
server.after_stop{
services.stop_server
}
}
The differences from the previous example is as follows.
code | description |
---|---|
Riser::DRbServices.new(0) |
setting the number of processes to 0 makes local process call without starting the dRuby server processes. |
server.process_num = 0 |
since local process call fails if it is a multi-process server, set it to single process multi-thread server. |
In this example, there are no dRuby server processes and there is only
1 server process. Looking at the process of this example with
pstree
command is as follows.
$ pstree -ap
...
| `-bash,23355
| `-ruby,3854 local_services.rb
| `-ruby,3855 local_services.rb
| |-{ruby},3856
| |-{ruby},3857
| |-{ruby},3858
| |-{ruby},3859
| `-{ruby},3860
...
The result of the web service always returns the same process id.
$ curl http://localhost:8000/any
pid: 3855
$ curl http://localhost:8000/any
pid: 3855
$ curl http://localhost:8000/any
pid: 3855
$ curl http://localhost:8000/single
pid: 3855
$ curl http://localhost:8000/single
pid: 3855
$ curl http://localhost:8000/single
pid: 3855
$ curl http://localhost:8000/sticky
key: default
pid: 3855
$ curl http://localhost:8000/sticky
key: default
pid: 3855
$ curl http://localhost:8000/sticky?foo
key: foo
pid: 3855
$ curl http://localhost:8000/sticky?foo
key: foo
pid: 3855
$ curl http://localhost:8000/sticky?bar
key: bar
pid: 3855
$ curl http://localhost:8000/sticky?bar
key: bar
pid: 3855
dRuby Services Callbacks
The object of Riser::DRbServices
is able to register callbacks.
The list of callbacks is as follows.
callback | description |
---|---|
at_fork(service_name) {|service_front| ... } |
performed when dRuby server process starts with remote process call. ignored by local process call. |
preprocess(service_name) {|service_front| ... } |
performed before starting the server. |
postprocess(service_name) {|service_front| ... } |
performed after the server stop. |
dRuby Services and Resource
An example of running a pstore database in a single process without collision in a multi-process server is as follows.
require 'pstore'
require 'riser'
Riser::Daemon.start_daemon(daemonize: false,
daemon_name: 'simple_count',
listen_address: 'localhost:8000'
) {|server|
services = Riser::DRbServices.new(1)
services.add_single_process_service(:pstore, PStore.new('simple_count.pstore', true))
server.process_num = 2
server.before_start{|server_socket|
services.start_server
}
server.at_fork{
services.detach_server
}
server.preprocess{
services.start_client
}
server.dispatch{|socket|
if (line = socket.gets) then
method, _uri, _version = line.split
while (line = socket.gets)
line.strip.empty? and break
end
if (method == 'GET') then
socket << "HTTP/1.0 200 OK\r\n"
socket << "Content-Type: text/plain\r\n"
socket << "\r\n"
services.get_service(:pstore).transaction do |pstore|
pstore[:count] ||= 0
pstore[:count] += 1
socket << 'count: ' << pstore[:count] << "\n"
end
end
end
}
server.after_stop{
services.stop_server
}
}
The result of the web service in this example is as follows.
$ curl http://localhost:8000/
count: 1
$ curl http://localhost:8000/
count: 2
$ curl http://localhost:8000/
count: 3
$ ls -l *.pstore
-rw-r--r-- 1 toki toki 13 Feb 5 15:21 simple_count.pstore
An example of using an undefined number of pstore is as follows. By
using Riser::ResourceSet
and sticky process pattern, you can create
a pstore object on access by each key.
require 'pstore'
require 'riser'
Riser::Daemon.start_daemon(daemonize: false,
daemon_name: 'simple_key_count',
listen_address: 'localhost:8000'
) {|server|
services = Riser::DRbServices.new(4)
services.add_sticky_process_service(:pstore,
Riser::ResourceSet.build{|builder|
builder.at_create{|key|
PStore.new("simple_key_count-#{key}.pstore", true)
}
builder.at_destroy{|pstore|
# nothing to do about `pstore'.
}
})
server.process_num = 2
server.before_start{|server_socket|
services.start_server
}
server.at_fork{
services.detach_server
}
server.preprocess{
services.start_client
}
server.dispatch{|socket|
if (line = socket.gets) then
method, uri, _version = line.split
while (line = socket.gets)
line.strip.empty? and break
end
if (method == 'GET') then
socket << "HTTP/1.0 200 OK\r\n"
socket << "Content-Type: text/plain\r\n"
socket << "\r\n"
_path, query = uri.split('?', 2)
key = query || 'default'
services.call_service(:pstore, key) {|pstore|
pstore.transaction do
pstore[:count] ||= 0
pstore[:count] += 1
socket << 'key: ' << key << "\n"
socket << 'count: ' << pstore[:count] << "\n"
end
}
end
end
}
server.after_stop{
services.stop_server
}
}
The result of the web service in this example is as follows.
$ curl http://localhost:8000/
key: default
count: 1
$ curl http://localhost:8000/
key: default
count: 2
$ curl http://localhost:8000/
key: default
count: 3
$ curl http://localhost:8000/?foo
key: foo
count: 1
$ curl http://localhost:8000/?foo
key: foo
count: 2
$ curl http://localhost:8000/?foo
key: foo
count: 3
$ curl http://localhost:8000/?bar
key: bar
count: 1
$ curl http://localhost:8000/?bar
key: bar
count: 2
$ curl http://localhost:8000/?bar
key: bar
count: 3
$ ls -l *.pstore
-rw-r--r-- 1 toki toki 13 Feb 5 16:16 simple_key_count-bar.pstore
-rw-r--r-- 1 toki toki 13 Feb 5 16:15 simple_key_count-default.pstore
-rw-r--r-- 1 toki toki 13 Feb 5 16:15 simple_key_count-foo.pstore
Development
After checking out the repo, run bin/setup
to install
dependencies. You can also run bin/console
for an interactive prompt
that will allow you to experiment.
To install this gem onto your local machine, run bundle exec rake install
. To release a new version, update the version number in
version.rb
, and then run bundle exec rake release
, which will
create a git tag for the version, push git commits and tags, and push
the .gem
file to rubygems.org.
Contributing
Bug reports and pull requests are welcome on GitHub at https://github.com/y10k/riser.
License
The gem is available as open source under the terms of the MIT License.