Raising the buffers is the default reaction to the nginx error upstream sent too big header while reading response header from upstream with php-fpm and fcgi, but it doesn't solve the problem, just makes it harder to hit.
Put this in a .php file somewhere:
<?php
for ($i = 0; $i<$_GET['iterations']; $i++)
error_log(str_pad("a", $_GET['size'], "a"));
echo "got here\n";Corrupt fpm error_log so it writes errors to stderr instead of a file:
bash-4.1# grep error_log /etc/php-fpm.d/foo.conf
php_admin_value[error_log] = /var/asdf/asdf/asdf.log
Restart fpm (you can figure this bit out).
Compare 502 errors with different error sizes/iterations...
Native configuration:
bash-4.1# for it in {30..200..3}; do for size in {100..250..3}; do echo "size=$size iterations=$it $(curl -sv "http://localhost/debug.php?size=$size&iterations=$it" 2>&1 | egrep '^< HTTP')"; done; done | grep 502 | head
size=121 iterations=30 < HTTP/1.1 502 Bad Gateway
size=109 iterations=33 < HTTP/1.1 502 Bad Gateway
size=232 iterations=33 < HTTP/1.1 502 Bad Gateway
size=241 iterations=48 < HTTP/1.1 502 Bad Gateway
size=145 iterations=51 < HTTP/1.1 502 Bad Gateway
size=226 iterations=51 < HTTP/1.1 502 Bad Gateway
size=190 iterations=60 < HTTP/1.1 502 Bad Gateway
size=115 iterations=63 < HTTP/1.1 502 Bad Gateway
size=109 iterations=66 < HTTP/1.1 502 Bad Gateway
size=163 iterations=69 < HTTP/1.1 502 Bad Gateway
[... there would be more here, but I piped through head ...]
fastcgi_buffers 16 16k; fastcgi_buffer_size 32k;:
bash-4.1# for it in {30..200..3}; do for size in {100..250..3}; do echo "size=$size iterations=$it $(curl -sv "http://localhost/debug.php?size=$size&iterations=$it" 2>&1 | egrep '^< HTTP')"; done; done | grep 502 | head
size=223 iterations=69 < HTTP/1.1 502 Bad Gateway
size=184 iterations=165 < HTTP/1.1 502 Bad Gateway
size=151 iterations=198 < HTTP/1.1 502 Bad Gateway