[MathWorks] Versioning Infrastructure

MERGE
2013
THE
PERFORCE
CONFERENCE

SAN
FRANCISCO
•
APRIL
24−26

Perforce White Paper
The
Perforce
Software
Configuration
Management

system
is
powerful
and
flexible.

Managing
a
powerful
SCM
requires
certain

infrastructure.

Changes
to
the
infrastructure
itself
can
become

challenging
and
destabilize
the
whole
system.

Using
the
flexibility
of
Perforce
can
help
manage
and

monitor
a
complex
infrastructure.

We
hope
our
experience
will
help
other
Perforce

administrators
plan
their
growing
systems.

Versioning Infrastructure
Automation and Scalability.
Administration Tips and Tricks.
Michael Mirman, MathWorks, Inc.)


2

Perforce at MathWorks
The company’s source code is the company’s treasure. Perforce is the system that protects it.
As of this writing, MathWorks has one main production server, about 700 users, and several
million archive files.
We deploy almost all possible triggers and several daemons and partially mirror our own bug
database into Perforce.
In addition to P4, P4V, P4Perl, P4Java, Emacs, and P4Eclipse interfaces, we support a
modified version of P4DB.1

1

There
are
several
versions
of
P4DB
publicly
available.

Our
version
is
close
to
the
version
at
http://www.releng.com/downloads3/index.html


3

Architecture Overview
We use multiple proxies, p4brokers, and multiple replicas for different purposes (see Figure 1).
Figure 1: Perforce architecture at Mathworks
Details:
1. Master, replica-1, replica-2 share the archive partition on an NFS disk. Replicas use “p4
pull -j” and replicate only meta-data. These replicas mount the archive as a read-only
disk.
2. Replica-3 is a separate full replica, using “p4 replicate” for replication. It is used to
checkpoint the database daily, as well as for other maintenance processes. In an
unlikely case of other replicas being down, this replica is a backup for p4broker-1.
3. P4broker-1 redirects some of the load to the “best” replica. If no replicas are available,
all queries are sent to the master.
4. If the master is unavailable, p4broker-1 redirects all the read-only requests to a replica
and produces a meaningful error message for other requests.
5. P4broker-2 serves one of the web interfaces to our CI system. All queries from that
system are read-only, and they are all directed to a standby replica by default. If that
Read-‐only

Proxy
server

p4broker-‐1
(H/A
VM)

master
replica-‐1
replica-‐2;
warm
standby
replica-‐3

p4broker-‐2
(H/A
VM)

Proxy
server
Proxy
server
Proxy
server
Proxy
server
Proxy
server

Read-‐only


4

replica is unavailable, another one is used. The master is used only if no replicas are
available.
How We Describe Our Infrastructure
Our infrastructure is described by an xml configuration file, where every server has certain
characteristics. In addition, we describe all “services” that must be started during the boot
process. For example, we may have this description in the file:
<variables>
<set>brokerhost=p4brokertest</set>
</variables>
<server>
<type>Primary</type>
<hostname>$brokerhost</hostname>
<port>1666</port>
<top>/local</top>
<start>
[/local/perforce/bin/p4broker -d -c
/local/perforce/config/broker.testconf]
</start>
<init_script>
/etc/init.d/p4broker.test -> /local/perforce/scripts/p4d.test
</init_script>
</server>
We have two files describing the infrastructure: one for the production stack and the other for
the test stack.
Both files are handled by a special and separate server. These files are the only files on that
server, so we don’t need a license for it.
The server has one trigger installed: change-commit. Every time one of those two files gets
updated, the trigger “pushes” the new file to all essential hosts of the infrastructure. (We will say
more about how we push files to different hosts later.)
Examples of How We Use Configuration Files Describing
Infrastructure
Example 1: Boot scripts and controlled failover.
We keep in /etc/init.d links to actual scripts. For example:
/etc/init.d/p4broker.prod -> /local/perforce/scripts/p4broker.prod*


5

Actual scripts read the configuration file, determine what they need to start on the current host,
and proceed accordingly.
Every Perforce host has a single boot script, and there are very few different scripts, which we
use as boot scripts (code used on p4d servers is slightly different from the code on p4p or
p4broker servers).
Say we are performing a controlled failover: the server that was a master will become a standby
and vice versa.
The old configuration may say:
<variables>
<set>master=p4test-00</set>
<set>standby=p4test-01</set>
</variables>
<server>
<type>Primary</type>
<hostname>$master</hostname>
<port>1666</port>
<top>/export</top>
<start>
[$bin/p4d -d -p 1666 -In MasterTest -r $dbroot/1666 -L
$logroot/1666/p4d.log –J $jnlroot/1666/journal]
</start>
<init_script>
/etc/init.d/p4d.current -> $dbroot/scripts/p4d.test
</init_script>
</server>
<server>
<type>Secondary</type>
<hostname>$standby</hostname>
<port>1666</port>
<top>/export</top>
<start>
[$bin/p4d -d -p 1666 -In StandbyTest -r $dbroot/1666 -L
$logroot/1666/p4d.log -J $jnlroot/1666/journal]
[$bin/p4 -p $production:1666 replicate -s $jnlroot/1666/replica.state
-J $jnlroot/1666/journal $dbroot/1666/scripts/p4admin_replicate -v -
port 1666 -srchost $production -srctop /export/data/perforce/1666]
</start>
<init_script>
/etc/init.d/p4d.current -> $dbroot/scripts/p4d.test
</init_script>
</server>


6

To perform the failover, all we have to do is flip the settings of master and standby:
<set>master=p4test-01</set>
<set>standby=p4test-00</set>
and then restart both servers.2
Example 2: Cron jobs.
Because some of our hosts change roles periodically, it is convenient to keep the crontabs
identical, so we do not have to change them every time the roles of the hosts change.
So some of our cron jobs read the configuration file, determine the role of the current host, and
can perform different functions depending on the role.
For example, production hosts have these cron lines:
perforce-03:17 01 * * * /local/perforce/scripts/cronjobs/p0X.cron
This script performs all daily maintenance on all the hosts, but replicas run verifications, which
we don’t run on the master for performance reasons.
Updating Workspaces Around the World
Once in awhile, we get a request to publish automatically a certain area of a Perforce repository
to make it available for users. Typical example: a small group of users update certain
documentation, which is read by a larger number of users from a shared location.
This is a very familiar task for Perforce administrators, and it can be easily accomplished by a
change-commit trigger.
Our First Implementation
We define a Perforce client with the directory to be updated as its root.
Our change-commit trigger changes the current working directory to the workspace and syncs
the workspace.
The immediate first problem with this solution is that most of the shared locations to be updated
are on the network, and the Perforce server does not mount network locations.

2

In
reality,
the
failover
has
several
steps.
In
particular,
first,
we
check
that
the
replica
is
in
sync
with
the
master.

Only
then
do
we
reconfigure
and
restart
the
broker
so
queries
will
be
correctly
directed.


7

A solution to this problem is to dedicate a machine that will mount network locations and will
allow SSH connection from the Perforce server.
For example, we have this line in the triggers table:
doc01 change-commit //team/one/main/documentation/...
"triggers/updateRefTree -p perforce:1666 -c one.internal"
This trigger runs an ssh to one of our hosts, where the workspace is mounted, and syncs the
workspace.
The user’s “p4 submit” command does not finish until the trigger exits, so the user knows that
the update succeeded before getting the command prompt back.
This procedure worked well until we received a request to update an area with the same path,
but local to every location around the world. It involved more than 20 locations, each with a
dedicated host ready for an ssh connection to update a local area. We had to solve several
problems:
• An ssh to a remote location can take a noticeably long time. Even if we run all ssh’s in
parallel (which may create a spike in processes on the server), they will still take more
time than an average user is willing to tolerate while waiting for “p4 submit” to complete.
• It’s not unusual for a remote location not to respond and hang or time out. Hung ssh’s on
the server will be accumulating and require periodic cleaning. Unfulfilled requests to
sync will cause remote workspaces to be out of sync.
Our Second Implementation: Pull, Don’t Push
Our alternative solution used counters as semaphores to indicate the need to sync.
The change-commit trigger sets one counter per client to be synced. The name of the counter is
mwsync.hostname.workspace_name
The value is the time when the trigger ran.
Every infrastructure host that is used to update some workspaces runs a cron job every five
minutes. The cron job checks with the master whether there is any pending work for the local
host. If so, it updates the set workspace and removes the counter.
If a remote host is temporarily down, it will eventually be brought back up, run that cron, and
update the workspace.
It is easy to monitor, too: we check all the mwsync.* counters, and if the counter value is
sufficiently old, we send email to Perforce administrators that a particular host requires
attention.
A simpler implementation is to run a cron job every N minutes on every client machine and sync
all necessary workspaces. However, monitoring in that case becomes more complicated. Using


8

counters with time stamps as the value makes monitoring (and alerting administrators) a trivial
task.
This implementation has been working very well with many different workspaces.
Interestingly enough, the need to maintain many crontabs was also a task worth mentioning.
Maintenance of Crontabs
Before receiving a request to update workspaces around the world, our Perforce infrastructure
contained about 20 hosts. Almost all of them had some cron jobs running regularly by the
special “service” user.
When we had to change a crontab, we would do what any UNIX administrator would do: rsh or
ssh as the service user to the machine, run “crontab –e”, and modify the crontab. A number of
people could log in as that service user, and we had no history of who modified the crontab,
when, and why. All we could do was to save daily the version we had. The result was
something like this:
00 05 * * * cd somedir; crontab -l > `hostname`; ci -l `hostname` >
/dev/null 2>&1
Now we had to approximately double the number of the Perforce hosts by adding new hosts
around the world. The thought that we would sometimes have to ssh to 40+ machines to modify
crontabs was not appealing.
Our Solution: Put Crontabs Under Perforce Control and Set a Change-
commit Trigger to Update the Crontab
Here is an example of a trigger line:
cron change-commit //depot/crontabs/... "triggers/install.crontab -p
%serverport% -c %changelist%"
The essence of the trigger is to run
p4 print -q //depot/crontabs/somefile | ssh host crontab -
We currently use the “push” method to install crontabs on our Perforce infrastructure hosts. If it
becomes too slow, we’ll switch to the “pull” method described earlier.
For a push method, a single network glitch may prevent us from updating the crontab. We
wanted to ensure that we eventually get the right one. So we have one cron job on each of
those hosts that runs once a day and updates the crontab this way:
01 00 * * * export HOST=`hostname`; /local/perforce/bin/p4 -p perforce:1666 print -q
//depot/crontabs/batscm/$HOST > /tmp/crontab.$$; [ -s /tmp/crontab.$$ ] && crontab
/tmp/crontab.$$; rm /tmp/crontab.$$


9

After all this was set up, if we need to change many crontabs at the same time (say, moving
maintenance tasks from 1 a.m. to 3 a.m.), it might be a simple sequence of commands like this:
p4 edit *
perl –pi.bak –e ‘s/^00 01/00 03/’ *
p4 submit –d “moving crons from 1am to 3am”
Conclusion
• The more stable your infrastructure is, the better you sleep at night. Use Perforce power
to increase stability of the infrastructure.
• Use a meta language to describe the infrastructure, so changes to the infrastructure are
as simple as changing the files that define it.
• If your servers change roles periodically (as in the failover scenario), keeping crontabs
on your servers nearly identical simplifies maintenance.
• Keep files describing the main infrastructure on a separate server to avoid bootstrapping
issues.
• Use change-commit triggers to update different pieces of the infrastructure (crontabs, for
example).
• Monitor whether change-commit scripts work and notify administrators in case of a long
pending update.

[MathWorks] Versioning Infrastructure

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to [MathWorks] Versioning Infrastructure (20)

More from Perforce (20)

Recently uploaded (20)

[MathWorks] Versioning Infrastructure