HAProxy 1.6.3-1ubuntu0.1统计时间异常及单位疑问咨询

Great questions—let’s unpack these two issues based on how HAProxy 1.6 handles request timing metrics.

1. Could ttime be in microseconds instead of milliseconds? #

Short answer: No, the units are definitely milliseconds in HAProxy 1.6’s stats. The official documentation confirms that qtime, ctime, rtime, and ttime all use millisecond granularity.

Your qtime, ctime, and rtime showing 0 makes sense here—those values are rounded down to the nearest millisecond. If each of those stages (queue wait, backend connection setup, server response) took less than 1ms, they’ll display as 0.

That said, a 2704ms (≈2.7s) ttime is indeed way outside your expected 0-10ms range. This suggests the delay is happening somewhere not captured by the other three metrics (we’ll get to that next).

2. Why is ttime so different from qtime + ctime + rtime? #

The key here is understanding what each metric actually measures:

• qtime: Time spent waiting in the queue before being assigned to a backend server
• ctime: Time taken to establish a connection to the backend server
• rtime: Time from when the backend connection is established to when HAProxy receives the last byte of the server’s response
• ttime: Total time from when HAProxy accepts the client connection to when it sends the last byte of the response back to the client

ttime includes far more than just the sum of the other three metrics. The missing time could come from:

• Slow client reception: If the client is slow to receive data (e.g., poor network, throttled client), the time HAProxy spends sending the response to the client counts towards ttime but isn’t tracked in the other metrics.
• HAProxy internal processing: Time spent parsing headers, rewriting requests/responses, or logging can add up, especially if you have complex configurations.
• Retries or redirects: If HAProxy had to retry the request on a different backend server, that extra time is included in ttime but not the individual stage metrics.
• Kept-alive connection overhead: If the connection was reused for multiple requests, idle time between requests might skew ttime for the final request in the sequence.

Troubleshooting Tips #

To get to the bottom of the high ttime:

• Check backend server logs to confirm how quickly the server actually processed the request. If server logs show a fast response, the delay is likely between HAProxy and the client.
• Inspect network latency between HAProxy and the client—tools like tcpdump or HAProxy’s detailed logging can help here.
• Review your HAProxy timeout settings (e.g., timeout client, timeout server) to ensure connections aren’t hanging unnecessarily.
• Enable detailed request logging with custom log-format strings that include timestamps for each stage of the request lifecycle. This will give you a granular breakdown of where the time is being spent.

内容的提问来源于stack exchange，提问作者clay