NAME
rq v2.3.2
SYNOPSIS
rq (queue | export RQ_Q=q) mode [mode_args]* [options]*
URIS
http://raa.ruby-lang.org/project/rq/
Linux Clustering with Ruby Queue: Small Is Beautiful | Linux Journal
DESCRIPTION
ruby queue (rq) is a zero-admin zero-configuration tool used to
create
instant linux clusters. rq requires only a central nfs filesystem in
order
to manage a simple sqlite database as a distributed priority work
queue.
this simple design allows researchers to install and configure, in
only a
few minutes and without root privileges, a robust linux cluster
capable of
distributing processes to many nodes - bringing dozens of powerful
cpus to
their knees with a single blow. clearly this software should be kept
out of
the hands of free radicals, seti enthusiasts, and one mr. j safran.
the central concept of rq is that n nodes work in isolation to pull
jobs
from an centrally mounted nfs priority work queue in a synchronized
fashion.
the nodes have absolutely no knowledge of each other and all
communication
is done via the queue meaning that, so long as the queue is available
via
nfs and a single node is running jobs from it, the system will
continue to
process jobs. there is no centralized process whatsoever - all nodes
work
to take jobs from the queue and run them as fast as possible. this
creates
a system which load balances automatically and is robust in face of
node
failures.
although the rq system is simple in it’s design it features powerful
functionality such as priority management, predicate and sql query ,
compact
streaming command-line processing, programmable api, hot-backup, and
input/capture of the stdin/stdout/stderr io streams of remote jobs.
to date
rq has had no reported runtime failures and is in operation at dozens
of
research centers around the world.
INVOCATION
the first argument to any rq command is the always the name of the
queue
while the second is the mode of operation. the queue name may be
omitted
if, and only if, the environment variable RQ_Q has been set to
contain the
absolute path of target queue.
for instance, the command
~ > rq queue list
is equivalent to
~ > export RQ_Q=queue
~ > rq list
this facility can be used to create aliases for several queues, for
example,
a .bashrc containing
alias MYQ="RQ_Q=/path/to/myq rq"
alias MYQ2="RQ_Q=/path/to/myq2 rq"
would allow syntax like
MYQ2 submit < joblist
MODES
rq operates in modes create, submit, resubmit, list, status, delete,
update,
query, execute, configure, snapshot, lock, backup, rotate, feed,
recover,
ioview, and help, and a few others. the meaning of ‘mode_args’ will
naturally change depending on the mode of operation.
the following mode abbreviations exist, note that not all modes have
abbreviations
c => create
s => submit
r => resubmit
l => list
ls => list
t => status
d => delete
rm => delete
u => update
q => query
e => execute
C => configure
S => snapshot
L => lock
b => backup
R => rotate
f => feed
io => ioview
h => help
create, c :
creates a queue. the queue must be located on an nfs mounted file
system
visible from all nodes intended to run jobs from it. nfs locking
must be
functional on this file system.
examples :
0) to create a queue
~ > rq /path/to/nfs/mounted/q create
or, using the abbreviation
~ > rq /path/to/nfs/mounted/q c
submit, s :
submit jobs to a queue to be proccesed by some feeding node. any
'mode_args' are taken as the command to run. note that 'mode_args'
are
subject to shell expansion - if you don’t understand what this
means do
not use this feature and pass jobs on stdin.
when running in submit mode a file may by specified as a list of
commands
to run using the ‘–infile, -i’ option. this file is taken to be a
newline separated list of commands to submit, blank lines and
comments (#)
are allowed. if submitting a large number of jobs the input file
method
is MUCH, more efficient. if no commands are specified on the
command line
rq automatically reads them from stdin. yaml formatted files are
also
allowed as input (http://www.yaml.org/) - note that the output of
nearly
all rq commands is valid yaml and may, therefore, be piped as input
into
the submit command. the leading ‘—’ of yaml file may not be
omitted.
when submitting the '--priority, -p' option can be used here to
determine
the priority of jobs. priorities may be any whole number including
negative ones - zero is the default. note that submission of a
high
priority job will NOT supplant a currently running low priority
job, but
higher priority jobs WILL always migrate above lower priority jobs
in the
queue in order that they be run as soon as possible. constant
submission
of high priority jobs may create a starvation situation whereby low
priority jobs are never allowed to run. avoiding this situation is
the
responsibility of the user. the only guaruntee rq makes regarding
job
execution is that jobs are executed in an ‘oldest-highest-priority’
order
and that running jobs are never supplanted. jobs submitted with
the
‘–stage’ option will not be eligible to be run by any node and
will
remain in a ‘holding’ state until updated (see update mode) into
the
‘pending’ mode, this option allows jobs to entered, or ‘staged’, in
the
queue and then made candidates for running at a later date.
rq allows the stdin of commands to be provided and also captures
the
stdout and stderr of any job run (of course standard shell
redirects may
be used as well) and all three will be stored in a directory
relative the
the queue itself. the stdin/stdout/stderr files are stored by job
id and
there location (though relative to the queue) is shown in the
output of
‘list’ (see docs for list).
examples :
0) submit the job ls to run on some feeding host
~ > rq q s ls
1) submit the job ls to run on some feeding host, at priority 9
~ > rq -p9 q s ls
2) submit a list of jobs from file. note the '-' used to specify
reading jobs from stdin
~ > cat joblist
job1.sh
job2.sh
job2.sh
~ > rq q submit --infile=joblist
3) submit a joblist on stdin
~ > cat joblist | rq q submit -
or
~ > rq q submit - <joblist
4) submit cat as a job, providing the stdin for cat from the file
cat.in
~ > rq q submit cat --stdin=cat.in
5) submit cat as a job, providing the stdin for the cat job on
stdin
~ > cat cat.in | rq q submit cat --stdin=-
or
~ > rq q submit cat --stdin=- <cat.in
6) submit 42 priority 9 jobs from a command file, marking them as
'important' using the '--tag, -t' option.
~ > wc -l cmdfile
42
~ > rq -p9 -timportant q s < cmdfile
6) re-submit all the 'important' jobs (see 'query' section below)
~ > rq q query tag=important | rq q s -
8) re-submit all jobs which are already finished (see 'list'
section
below)
~ > rq q l f | rq q s
9) stage the job wont_run_yet to the queue in a 'holding' state.
no
feeder will run this job until it’s state is upgraded to
‘pending’
~ > rq q s --stage wont_run_yet
ioview, io :
as shown in the description for submit, a job maybe be provided
stdin
during job submission. the stdout and stderr of the job are also
captured
as the job is run. all three streams are captured in files located
relative to the queue. so, if one has submitted a job, and it’s
jid was
shown to be 42, by using something like
~ > rq /path/to/q submit myjob --stdin=myjob.in
---
-
jid : 42
priority : 0
...
stdin : stdin/42
stdout : stdout/42
stderr : stderr/42
...
command : myjob
the stdin file will exists as soon as the job is submitted and the
others
will exist once the job has begun running. note that these paths
are
shown relative to the queue. in this case the actual paths would
be
/path/to/q/stdin/42
/path/to/q/stdout/42
/path/to/q/stderr/42
but, since our queue is nfs mounted the /path/to/q may or may not
be the
same on every host. thus the path is a relative one. this can
make it
anoying to view these files, but rq assists here with the ioview
command.
the ioview command spawns an external editor to view all three
files.
it’s use is quite simple
examples :
0) view the stdin/stdout/stderr of job id 42
~ > rq ioview 42
by default this will open up all three files in vim. the editor
command
can be specified using the ‘–editor’ option or the ENV var
RQ_EDITOR.
the default value is ‘vim -R -o’ which allows all three files to be
opened
in a single window.
resubmit, r :
resubmit jobs back to a queue to be proccesed by a feeding node.
resubmit
is essentially equivalent to submitting a job that is already in
the queue
as a new job and then deleting the original job except that using
resubmit
is atomic and, therefore, safer and more efficient. resubmission
respects
any previous stdin provided for job input. read docs for delete
and
submit for more info.
examples :
0) resubmit job 42 to the queue
~> rq q resubmit 42
1) resubmit all failed jobs
~> rq q query exit_status!=0 | rq q resubmit -
2) resubmit job 4242 with different stdin
~ rq q resubmit 4242 --stdin=new_stdin.in
list, l, ls :
list mode lists jobs of a certain state or job id. state may be
one of
pending, holding, running, finished, dead, or all. any ‘mode_args’
that
are numbers are taken to be job id’s to list.
states may be abbreviated to uniqueness, therefore the following
shortcuts
apply :
p => pending
h => holding
r => running
f => finished
d => dead
a => all
examples :
0) show everything in q
~ > rq q list all
or
~ > rq q l all
or
~ > export RQ_Q=q
~ > rq l
1) show q's pending jobs
~ > rq q list pending
2) show q's running jobs
~ > rq q list running
3) show q's finished jobs
~ > rq q list finished
4) show job id 42
~ > rq q l 42
5) show q's holding jobs
~ > rq q list holding
status, t :
status mode shows the global state the queue and statistics on it's
the
cluster’s performance. there are no ‘mode_args’. the meaning of
each
state is as follows:
pending => no feeder has yet taken this job
holding => a hold has been placed on this job, thus no feeder
will start
it
running => a feeder has taken this job
finished => a feeder has finished this job
dead => rq died while running a job, has restarted, and moved
this job to the dead state
note that rq cannot move jobs into the dead state unless it has
been
restarted. this is because no node has any knowledge of other
nodes and
cannot possibly know if a job was started on a node that died, or
is
simply taking a very long time. only the node that dies, upon
restart,
can determine that is has jobs that ‘were started before it
started’ and
move these jobs into the dead state. normally only a machine crash
would
cause a job to be placed into the dead state. dead jobs are never
automatically restarted, this is the responsibility of an operator.
examples :
0) show q's status
~ > rq q t
delete, d :
delete combinations of pending, holding, finished, dead, or jobs
specified
by jid. the delete mode is capable of parsing the output of list
and
query modes, making it possible to create custom filters to delete
jobs
meeting very specific conditions.
'mode_args' are the same as for list.
note that it is NOT possible to delete a running job. rq has a
decentralized architechture which means that compute nodes are
completely
independant of one another; an extension is that there is no way to
communicate the deletion of a running job from the queue the the
node
actually running that job. it is not an error to force a job to
die
prematurely using a facility such as an ssh command spawned on the
remote
host to kill it. once a job has been noted to have finished,
whatever the
exit status, it can be deleted from the queue.
examples :
0) delete all pending, finished, and dead jobs from a queue
~ > rq q d all
1) delete all pending jobs from a queue
~ > rq q d p
2) delete all finished jobs from a queue
~ > rq q d f
3) delete jobs via hand crafted filter program
~ > rq q list | yaml_filter_prog | rq q d -
an example ruby filter program (you have to love this)
~ > cat yaml_filter_prog
require 'yaml'
joblist = YAML::load STDIN
y joblist.select{|job| job['command'] =~ /bombing_program/}
this program reads the list of jobs (yaml) from stdin and then
dumps
only those jobs whose command matches ‘bombing_program’, which
is
subsequently piped to the delete command.
update, u :
update assumes all leading arguments are jids to update with
subsequent
key=value pairs. currently only the ‘command’, ‘priority’, and
‘tag’
fields of pending jobs can be generically updated and the ‘state’
field
may be toggled between pending and holding.
examples:
0) update the priority of job 42
~ > rq q update 42 priority=7
1) update the priority of all pending jobs
~ > rq q update pending priority=7
2) query jobs with a command matching 'foobar' and update their
command
to be ‘barfoo’
~ > rq q q "command like '%foobar%'" |\
rq q u command=barfoo
3) place a hold on jid 2
~ > rq q u 2 state=holding
4) place a hold on all jobs with tag=disk_filler
~ > rq q q tag=disk_filler | rq q u state=holding -
5) remove the hold on jid 2
~ > rq q u 2 state=pending
query, q :
query exposes the database more directly the user, evaluating the
where
clause specified on the command line (or read from stdin). this
feature
can be used to make a fine grained slection of jobs for reporting
or as
input into the delete command. you must have a basic understanding
of SQL
syntax to use this feature, but it is fairly intuitive in this
limited
capacity.
examples:
0) show all jobs submitted within a specific 10 minute range
~ > a='2004-06-29 22:51:00'
~ > b='2004-06-29 22:51:10'
~ > rq q query "started >= '$a' and started < '$b'"
1) shell quoting can be tricky here so input on stdin is also
allowed to
avoid shell expansion
~ > cat constraints.txt
started >= '2004-06-29 22:51:00' and
started < '2004-06-29 22:51:10'
~ > rq q query < contraints.txt
or (same thing)
~ > cat contraints.txt| rq q query -
2) this query output might then be used to delete those jobs
~ > cat contraints.txt | rq q q - | rq q d -
3) show all jobs which are either finished or dead
~ > rq q q "state='finished' or state='dead'"
4) show all jobs which have non-zero exit status
~ > rq q query exit_status!=0
5) if you plan to query groups of jobs with some common feature
consider
using the ‘–tag, -t’ feature of the submit mode which allows a
user to
tag a job with a user defined string which can then be used to
easily
query that job group
~ > rq q submit --tag=my_jobs - < joblist
~ > rq q query tag=my_jobs
6) in general all but numbers will need to be surrounded by
single
quotes unless the query is a ‘simple’ one. a simple query is a
query
with no boolean operators, not quotes, and where every part of it
looks
like
key op value
with ** NO SPACES ** between key, op, and value. if, and only
if,
the query is ‘simple’ rq will contruct the where clause
appropriately. the operators accepted, and their meanings,
are
= : equivalence : sql =
=~ : matches : sql like
!~ : not matches : sql not like
match, in the context is ** NOT ** a regular expression but a
sql
style string match. about all you need to know about sql
matches is
that the ‘%’ char matches anything. multiple simple queries
will be
joined with boolean ‘and’
this sounds confusing - it isn't. here are some examples of
simple
queries
6.a)
query :
rq q query tag=important
where_clause :
"( tag = 'important' )"
6.b)
query :
rq q q priority=6 restartable=true
where_clause :
"( priority = 6 ) and ( restartable = 'true' )"
6.c)
query :
rq q q command=~%bombing_job% runner=~%node_1%
where_clause :
"( command like '%bombing_job%') and (runner like
‘%node_1%’)"
execute, e :
execute mode is to be used by expert users with a knowledge of sql
syntax
only. it follows the locking protocol used by rq and then allows
the user
to execute arbitrary sql on the queue. unlike query mode a write
lock on
the queue is obtained allowing a user to definitively shoot
themselves in
the foot. for details on a queue’s schema the file ‘db.schema’ in
the
queue directory should be examined.
examples :
0) list all jobs
~ > rq q execute 'select * from jobs'
configure, C :
this mode is not supported yet.
snapshot, p :
snapshot provides a means of taking a snapshot of the q. use this
feature
when many queries are going to be run; for example when attempting
to
figure out a complex pipeline command your test queries will not
compete
with the feeders for the queue’s lock. you should use this option
whenever possible to avoid lock competition.
examples:
0) take a snapshot using default snapshot naming, which is made
via the
basename of the q plus ‘.snapshot’
~ > rq /path/to/nfs/q snapshot
1) use this snapshot to chceck status
~ > rq ./q.snapshot status
2) use the snapshot to see what's running on which host
~ > rq ./q.snapshot list running | grep `hostname`
note that there is also a snapshot option - this option is not the
same as
the snapshot command. the option can be applied to ANY command. if
in
effect then that command will be run on a snapshot of the database
and the
snapshot then immediately deleted. this is really only useful if
one were
to need to run a command against a very heavily loaded queue and
did not
wish to wait to obtain the lock. eg.
0) get the status of a heavily loaded queue
~ > rq q t --snapshot
1) same as above
~ > rq q t -s
** IMPORTANT **
a really great way to hang all processing in your queue is to do
this
rq q list | less
and then leave for the night. you hold a read lock you won't
release
until less dies. this is what snapshot is made for! use it like
rq q list -s | less
now you've taken a snapshot of the queue to list so your locks
affect no
one.
lock, L :
lock the queue and then execute an arbitrary shell command. lock
mode
uses the queue’s locking protocol to safely obtain a lock of the
specified
type and execute a command on the user’s behalf. lock type must be
one of
(r)ead | (sh)ared | (w)rite | (ex)clusive
examples :
0) get a read lock on the queue and make a backup
~ > rq q L read -- cp -r q q.bak
(the '--' is needed to tell rq to stop parsing command line
options which allows the '-r' to be passed to the 'cp'
command)
** IMPORTANT **
this is another fantastic way to freeze your queue - use with
care!
backup, b :
backup mode is exactly the same as getting a read lock on the queue
and
making a copy of it. this mode is provided as a convenience.
0) make a backup of the queue using default naming ( qname +
timestamp +
.bak )
~ > rq q b
1) make a backup of the queue as 'q.bak'
~ > rq q b q.bak
rotate, r :
rotate mode is conceptually similar to log rolling. normally the
list of
finished jobs will grow without bound in a queue unless they are
manually
deleted. rotation is a method of trimming finished jobs from a
queue
without deleting them. the method used is that the queue is copied
to a
‘rotation’; all jobs that are dead or finished are deleted from the
original queue and all pending and running jobs are deleted from
the
rotation. in this way the rotation becomes a record of the queue’s
finished and dead jobs at the time the rotation was made.
0) rotate a queue using default rotation name
~ > rq q rotate
1) rotate a queue naming the rotation
~ > rq q rotate q.rotation
2) a crontab entry like this could be used to rotate a queue
daily
59 23 * * * rq q rotate `date +q.%Y%m%d`
feed, f :
take jobs from the queue and run them on behalf of the submitter as
quickly as possible. jobs are taken from the queue in an 'oldest
highest
priority’ first order.
feeders can be run from any number of nodes allowing you to harness
the
CPU power of many nodes simoultaneously in order to more
effectively
clobber your network, anoy your sysads, and set output raids on
fire.
the most useful method of feeding from a queue is to do so in
daemon mode
so that if the process loses it’s controling terminal it will not
exit
when you exit your terminal session. use the ‘–daemon, -d’ option
to
accomplish this. by default only one feeding process per host per
queue
is allowed to run at any given moment. because of this it is
acceptable
to start a feeder at some regular interval from a cron entry since,
if a
feeder is alreay running, the process will simply exit and
otherwise a new
feeder will be started. in this way you may keep feeder processing
running even acroess machine reboots without requiring sysad
intervention
to add an entry to the machine’s startup tasks.
examples :
0) feed from a queue verbosely for debugging purposes, using a
minimum
and maximum polling time of 2 and 4 respectively. you would
NEVER
specify polling times this brief except for debugging purposes!!!
~ > rq q feed -v4 --min_sleep=2 --max_sleep=4
1) same as above, but viewing the executed sql as it is sent to
the
database
~ > RQ_SQL_DEBUG=1 rq q feed -v4 --min_sleep=2 --max_sleep=4
2) feed from a queue in daemon mode - logging to
/home/ahoward/rq.log
~ > rq q feed --daemon -l/home/$USER/rq.log
log rolling in daemon mode is automatic so your logs should
never
need to be deleted to prevent disk overflow.
start :
the start mode is equivalent to running the feed mode except the
–daemon
is implied so the process instantly goes into the background.
also, if no
log (–log) is specified in start mode a default one is used. the
default
is /home/$USER/$BASENAME_OF_Q.log
examples :
0) start a daemon process feeding from q
~ > rq q start
1) use something like this sample crontab entry to keep a feeder
running
forever - it attempts to (re)start every fifteen minutes but
exits if
another process is already feeding. output is only created when
the
daemon is started so your mailbox will not fill up with this
crontab
entry:
#
# crontab.sample
#
*/15 * * * * /path/to/bin/rq /path/to/q start
and entry like this on every node in your cluster is all that's
needed
to keep your cluster going - even after a reboot.
shutdown :
tell a running feeder to finish any pending jobs and then to exit.
this
is equivalent to sending signal ‘SIGTERM’ to the process - this is
what
using ‘kill pid’ does by default.
examples :
0) stop a feeding process, if any, that is feeding from q. allow
all
jobs to be finished first.
~ > rq q shutdown
** VERY IMPORTANT **
if you are keeping your feeder alive with a crontab entry you'll
need to
comment it out before doing this or else it will simply
re-start!!!
stop :
tell any running feeder to stop NOW. this sends signal 'SIGKILL'
(-9) to
the feeder process. the same warning as for shutdown applies!!!
examples :
0) stop a feeding process, if any, that is feeding from q. allow
NO
jobs to be finished first - exit instantly.
~ > rq q stop
pid :
show the pid, if any, of the feeder on this host
~ > rq q feeder
---
pid : 3176
help, h :
this message
examples :
0) get this message
~> rq q help
or
~> rq help
NOTES
-
realize that your job is going to be running on a remote host and
this has
implications. paths, for example, should be absolute, not
relative.
specifically the submitted job script must be visible from all
hosts
currently feeding from a queue as must be the input and output
files/directories. -
jobs are currently run under the bash shell using the --login
option.
therefore any settings in your .bashrc will apply - specifically
your PATH
setting. you should not, however, rely on jobs running with any
given
environment. -
you need to consider CAREFULLY what the ramifications of having
multiple instances of your program all potentially running at the
same
time will be. for instance, it is beyond the scope of rq to ensure
multiple instances of a given program will not overwrite each
others
output files. coordination of programs is left entirely to the
user. -
the list of finished jobs will grow without bound unless you
sometimes
delete some (all) of them. the reason for this is that rq cannot
know
when the user has collected the exit_status of a given job, and so
keeps
this information in the queue forever until instructed to delete
it. if
you have collected the exit_status of you job(s) it is not an error
to
then delete that job from the finished list - the information is
kept for
your informational purposes only. in a production system it would
be
normal to periodically save, and then delete, all finished jobs. -
know that it is a VERY bad idea to spawn several dozen process all
reading/writing huge output files to a single NFS server. use this
paradigm instead- copy/move data from global input space to local disk
- process data
- move data on local disk to global output space
this, of course, applies to any nfs processing, not just those jobs
submitted to rqthe vsftp daemon is an excellent utility to have running on hosts
in your
cluster so anonymous ftp can be used to get/put data between any
two
hosts. -
know that nfs locking is very, very easy to break with firewalls
put in
place by overzealous system administrators. be postive not only
that nfs
locking works, but that lock recovery server/client crash or reboot
works
as well. http://nfs.sourceforge.net/ is the place to learn about
NFS. my
experience thus far is that there are ZERO properly configured NFS
installations in the world. please test yours. contact me for a
simple
script which can assist you. beer donations required as payment.
ENVIRONMENT
RQ_Q: set to the full path of nfs mounted queue
the queue argument to all commands may be omitted if, and only if,
the
environment variable ‘RQ_Q’ contains the full path to the q. eg.
~ > export RQ_Q=/full/path/to/my/q
this feature can save a considerable amount of typing for those
weak of
wrist.
DIAGNOSTICS
success : $? == 0
failure : $? != 0
CREDITS
- kim baugh : patient tester and design input
- jeff safran : the guy can break anything
- chris elvidge : made it possible
- trond myklebust : tons of help with nfs
- jamis buck : for writing the sqlite bindings for ruby
- _why : for writing yaml for ruby
- matz : for writing ruby
AUTHOR
[email protected]
BUGS
0 < bugno && bugno <= 42
reports to [email protected]
OPTIONS
–priority=priority, -p
modes <submit, resubmit> : set the job(s) priority - lowest(0)
…
highest(n) - (default 0)
–tag=tag, -t
modes <submit, resubmit> : set the job(s) user data tag
–runner=runner
modes <submit, resubmit> : set the job(s) required runner(s)
–restartable
modes <submit, resubmit> : set the job(s) to be restartable on
node
reboot
–stage
modes <submit, resubmit> : set the job(s) initial state to be
holding
(default pending)
–infile=infile, -i
modes <submit, resubmit> : infile
–stdin=[stdin], -s
modes <submit, resubmit, update> : stdin
–quiet, -q
modes <submit, resubmit, feed> : do not echo submitted jobs,
fail
silently if another process is already feeding
–daemon, -D
modes : spawn a daemon
–max_feed=max_feed
modes : the maximum number of concurrent jobs run
–retries=retries
modes : specify transaction retries
–min_sleep=min_sleep
modes : specify min sleep
–max_sleep=max_sleep
modes : specify max sleep
–snapshot, -s
operate on snapshot of queue
–editor=editor, -e
editor command capable of opening multiple files at once =
(default
ENV[“RQ_EDITOR”] || “vim -R -o”)
–verbosity=verbostiy, -v
0|fatal < 1|error < 2|warn < 3|info < 4|debug - (default info)
–log=path, -l
set log file - (default stderr)
–log_age=log_age
daily | weekly | monthly - what age will cause log rolling
(default
nil)
–log_size=log_size
size in bytes - what size will cause log rolling (default nil)
–help, -h
this message
–version
show version number
enjoy.
-a
some people, sweet and attractive, and strong and healthy, happen to die
young. they are masters in disguise teaching us about impermanence.
- h.h. the 14th dali lama