https://bz.apache.org/bugzilla/show_bug.cgi?id=65180
--- Comment #3 from Christopher Pereira <[email protected]> --- I have been seen exactly the same behaviour with mod_proxy since some years/versions ago including latests CentOS version (httpd-2.4.6-97.el7.centos.x86_64). We are proxying CouchDB with this config: AllowEncodedSlashes On ProxyPass /couchdb http://localhost:5984/ nocanon ProxyPassReverse /couchdb http://localhost:5984/ The problem ocurrs with CouchDB longpoll or continuous feed connections (AKA "Live replication" or "continuous replication" in PouchDB jargon), ie. those HTTP connections that stay alive. The proxied connections are listed as "W" (sending reply) by apache status and all at once (!) they change all to "G" (gracefully finishing) no matter what their SS time is (seconds since beginning of most recent request). I'm not sure what triggers it, but it happens many times in a day, so I can provide more information if you need. strace shows that the processes in "G" state continue reading and writing data: read(73, "1\r\n\n\r\n", 8000) = 6 read(73, 0x7fd2a9cd05b8, 8000) = -1 EAGAIN (Resource temporarily unavailable) writev(72, [{"\27\3\3\0\34\265\221\7\333\31s\324\204\20\7\323{\30\35h\311\236\253}\306H\274B\36>\274\375"..., 33}, {"\27\3\3\0\32\265\221\7\333\31s\324\205y\32\332o\220YgkIo_\270\336\0z\300\355n", 31}], 2) = 64 read(73, 0x7fd2a9cd05b8, 8000) = -1 EAGAIN (Resource temporarily unavailable) poll([{fd=73, events=POLLIN}], 1, 60000) = 1 ([{fd=73, revents=POLLIN}]) Here fd = 73 is the connection between Apache and CouchDB (LAN) and fd = 72 is the connection between Apache and the PouchDB client (WAN). After restarting CouchDB, all threads in "G" state become an open slot again (the same PID continues running). The problem is that at some point, all slots will be used and the server gets irresponsive. Increasing MaxClients may also crash the server because each process consumes like 60 MB. -- You are receiving this mail because: You are the assignee for the bug. --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
