cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ron Kuris (JIRA)" <>
Subject [jira] [Commented] (CASSANDRA-9536) The failure detector becomes more sensitive when the network is flakey
Date Tue, 02 Jun 2015 19:52:50 GMT


Ron Kuris commented on CASSANDRA-9536:

Some additional notes from my research on how this code works:

FailureDetector's report() is called each time the gossiper receives a message from that node.
The message should arrive every second (this is not tunable). The failure detector records
the time since the last message, up to 1000 samples, which will take about 16 minutes, 40
seconds. It does not record the time if the message was sent more than 2 seconds ago.

In the gossiper's status check, it calls FailureDetector's interpret() method. This calculates
"phi", which is the time
since the last message, divided by the average of the collected samples. It then divides by
a constant, PHI_FACTOR, which is about 0.434, then compares to phi_convict_threshold. If the
adjusted phi is greater than the threshold, the node is marked down.

If the average collected samples drops below 1, this increases phi.

I ran several additional tests, including changing the packet dropping time to 0-1 seconds,
which severely lowers the average and causes the node to flip flop UP/DOWN.

I'm very open to helping to fix this, but I'm afraid it might require a different implementation
of the FailureDetector. Perhaps I should make it pluggable, where we have a new implementation
of IFailureDetector? I'm open to suggestions.

> The failure detector becomes more sensitive when the network is flakey
> ----------------------------------------------------------------------
>                 Key: CASSANDRA-9536
>                 URL:
>             Project: Cassandra
>          Issue Type: Improvement
>          Components: Core
>            Reporter: Ron Kuris
>            Priority: Minor
> I added considerable instrumentation into the failure detector, and then blocked port
7000 for a random 5-6 second interval, then resumed traffic for the same amount of time, with
a script like:
> {code}while :
> do
>    iptables -A INPUT -p tcp --destination-port 7000 -j DROP
>         v=$[100 + (RANDOM % 100)]$[1000 + (RANDOM % 1000)]
>         s=5.${v:1:2}${v:4:3}
>         echo offline for $s
>         sleep $s
>         iptables -F
>         v=$[100 + (RANDOM % 100)]$[1000 + (RANDOM % 1000)]
>         s=5.${v:1:2}${v:4:3}
>         echo online for $s
>         sleep $s
> done{code}
> When I do this, I watch the values being reported to the FailureDetector. The median
actually goes down, as low as 850ms. The reason is that the very slow packets are not recorded
(they exceed MAX_INTERVAL_IN_NANO which is 2 seconds) and the retransmitted packets arrive
very quickly in succession, lowering the overall average.
> Once the average is lowered, the node becomes much more sensitive to shorter outages.
If you run this code for a while, the average drops down to 800ms or less, which means that
the node will go down 20% quicker than expected.

This message was sent by Atlassian JIRA

View raw message