| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
| |
In my patch for #3088 I made a erroneous assumption about the ruby exception
hierarchy and thus missed the fact that Timeout::error descends from both
SignalError and Interrupt. This is a partial reversion of the patch for #3088
to let these through so that more useful error messages can be produced.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Changing rescues from the default to Exception (to catch errors that don't
descend from StandardError) had the unintended consequence of catching (and
suppressing) SystemExit.
This patch restores the behavior of by reraising the exception.
Of the other exceptions that fall through the same crack (NoMemoryError,
SignalException, LoadError, Interrupt, NotImplementedError, and ScriptError)
this patch also reraises NoMemoryError, SignalException, and Interrupt in the
same way and leaves the rest captured.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Disables the b64_zlib_yaml format if zlib cannot be loaded.
I've added a --no-zlib to make it possible to test this on a single
machine, but it might also be useful if someone finds themselves failing
to connect to a server that doesn't have zlib installed.
FactHandler' format is still hard-coded to YAML rather than using
facts.class.default_format
Signed-off-by: Jesse Wolfe <jes5199@gmail.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
If setup code for a process depends on network connectivity
it needs to be protected with a rescue clause as much as the
main body of the process.
Further, Timeout exceptions aren't under StandardError and thus
aren't caught by an un-typed rescue clause. This doesn't matter
if we've morphed the exception, but will cause the program to
fail if we haven't.
There are many places where these concerns _might_ cause a problem
but in most cases they never will in practice; this patch addresses
the five cases where I have been able to confirm that it actually
can cause the client daemon to exit and two more where I suspect
(but can not prove) that it could.
This is an extension of the prior patch to cover additional cases
found by automated testing (repeated catalog runs with a 1% chance
of timeout forced on all timeout-bound operations, ~5000 runs).
The new cases recurred multiple times (>100 each) and in a final pass
with these corrected (~2500 runs) no additional cases were found.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
less often
This is not the right fix, but more a hackish workaround.
Since 0.25, the facts are transmitted as GET parameters when a
node asks for a catalog. Most proxies or webserver have a size limit
which is sometimes reached. In this case the request is denied
and the node can't get its catalog.
The idea is to compress facts (some non-scientific studies show a
57% fact size decrease for an average node) when transmitting
those when asking for a catalog.
Signed-off-by: Brice Figureau <brice-puppet@daysofwonder.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The problem is that URI.escape by default doesn't escape '+' (and
some other characters). But some web framework (at least webrick)
unescape the query string behind Puppet's back changing all '+'
to spaces corrupting facts containing '+' characters (like base64
encoded values).
The current fix makes sure we use CGI.escape for all query string
parameters. Indirection keys/path are still using URI escaping because
this part of the URI format shouldn't be handled like query string
parameters (otherwise '/' url separators are encoded which changes
the uri path).
Signed-off-by: Brice Figureau <brice-puppet@daysofwonder.com>
|
| |
|
|
|
|
|
| |
Previously any changed file got loaded; now we only try to
load files that are still present.
Signed-off-by: Luke Kanies <luke@madstop.com>
|
| |
|
|
| |
Signed-off-by: Luke Kanies <luke@madstop.com>
|
| |
|
|
|
|
|
|
| |
We weren't splitting on whitespace, which is necessary
since the settings don't support arrays but files
expect them.
Signed-off-by: Luke Kanies <luke@madstop.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This removes the requirement of shared fact caching
on the servers, since the server responding to the catalog
request will receive the facts as part of the request.
The facts are serialized as a parameter to the request,
rather than each being set as a separate request parameter.
This hard-codes yaml as the serialization format for the
facts, because I couldn't get marshal to work and it's just not
as big a deal for such a small amount of data.
Signed-off-by: Luke Kanies <luke@madstop.com>
|
| |
|
|
| |
semicolons
|
| |
|
|
|
|
|
|
| |
At this point, the server's behaviour is a bit undefined
if it tries to compile the catalog with no facts
locally. The next commits will fix that.
Signed-off-by: Luke Kanies <luke@madstop.com>
|
| |
|
|
|
|
|
| |
create() got deprecated and I apparently missed
changing this.
Signed-off-by: Luke Kanies <luke@madstop.com>
|
| |
|
|
|
|
|
|
| |
Also fixing the argument order while downloading
either of them. I had my Downloader.new
calls using the wrong argument order.
Signed-off-by: Luke Kanies <luke@madstop.com>
|
| |
|
|
|
|
|
| |
This will eventually be used by puppetrun, but
for now is just called by the old-school Runner handler.
Signed-off-by: Luke Kanies <luke@madstop.com>
|
|
|
Once I went to add runinterval support to the Agent class,
I realized it's really two classes: One that handles starting,
stopping, running, et al (still called Agent), and one that
handles downloading the catalog, running it, etc. (now
called Configurer).
This commit includes some additional code, but 95% of it is just moving code around.
Signed-off-by: Luke Kanies <luke@madstop.com>
|
