What do we mean by "top percentile" or TP based latency?

Question

When we discuss performance of a distributed system we use the terms tp50, tp90, tp99.99 TPS. Could someone explain what do we mean by those?

Solution

tp90 is a maximum time under which 90% of requests have been served. Imagine you have times:

10s
1000s
100s
2s

Calculating TP is very simple:

• sort all times in ascending order: [2s, 10s, 100s, 1000s]
• find latest item in portion you need to calculate. For TP50 it will ceil(4*.5)=2 requests. You need 2nd request. For TP90 it will be ceil(4*.9)=4. You need 4th request.
• get time for the item found above. TP50=10s. TP90=1000s

Comments

[sichinumi]

This doesn't seem to match with the statistical definition of a percentile. Instead of using a ceiling to find an index, you should be averaging the two closest indices. For example, in a 4-element list, TP50 is the average of the 2nd and 3rd elements, not just the 2nd element. (This is assuming TP50 means the 50th percentile.) The exact formula is i=(k/100)(n+1) to find your desired index, where k is your percentile and n is the number of elements in your list. If i is not a whole number, average the two nearest indices.

[Underoos]

I think it is maximum time. Not minimum time

[Sergey Romanovsky]

@Underoos you're right, I fixed the typo, thanks!