timeouts in doctests on OS X 10.9

[jhpalmieri]

There are frequent timeouts when doctesting on OS X with pexpect interfaces, in particular Singular. For example:

$ ./sage -tp src/doc/en/constructions

...
----------------------------------------------------------------------
sage -t src/doc/en/constructions/algebraic_geometry.rst  # Timed out
sage -t src/doc/en/constructions/polynomials.rst  # Timed out
----------------------------------------------------------------------

These became much more frequent by the changes in the Singular/pexpect interface from #15631 on two buildbots:

• OSX 10.9 (Oxford)
• Ubuntu 8.04 (UW)
• This happens mostly during parallel doctesting, and disappers when the system load is low

Other build bots do not seem to be affected, so its a likely pty kernel bug.

[dimpase]

even after reverting the corresponding commit from #15631 I still get lots of doctest timeouts.

[jhpalmieri]

I'm not suggesting this as a fix (because I don't know anything about the pexpect interface), but if I make this one change:

src/sage/interfaces/singular.py

@@ -378 +378 @@
         """
         prompt = '> '
         Expect.__init__(self,
-                        terminal_echo=False,
+                        terminal_echo=True,
                         name = 'singular',
                         prompt = prompt,
                         command = "Singular -t --ticks-per-sec 1000", #no tty and fine grained cputime()

then I don't get the timeouts any more. (I get other doctest failures because of this, but no timeouts.)

[dimpase]

in case, note #16260#comment:62

[vbraun]

I've debugged the issue on OSX 10.9 and when the hang happens

• Singular does print its header and prompt (captured with dtrace)
• Python never sees any output from the subprocess, the first select() in pexpect.read_nonblocking hangs forever.

One way to reduce the likelihood of the hang is to insert a sleep before first reading from the subproces (in expect._start)

[vbraun]

Sleeping 100ms when starting Singular reduces the failure rate in sage -t --long src/sage/interfaces/magma.py from about 1/10 to 1/1000 fwiw.

[vbraun]

The branch attached is obviously a no-op, and just moves the setecho(0) further down after we read the first prompt. But it seems to improve things in my limited testing.

---

New commits:

​2928b9a

Fix Singular pexpect timeouts on some platforms

[vbraun]

This seems to fix it, I haven't seen another failure in sage -t --long src/sage/interfaces/magma.py

[jhpalmieri]

Oh, very nice. This fixes it on two OS X 10.9 machines.