Two Problems With One Name
The stampede everybody knows is the concurrent one: a popular key expires and fifty requests arriving in the same millisecond all fetch it. Single flight fixes that, it has been measured three times already in this chapter, and it is half the subject.
The mechanisms that fix them are unrelated, and each does nothing for the other. Most discussions of stampede prevention describe only the left panel.
The cold start is over in a minute. What outlives it is that all of those entries now share an expiry schedule, and the refill repeats as a spike every time the time to live comes round.
The Echo Of A Deploy
With no jitter the busiest second issued 506 loads against a mean of 251. The cache was filled in a burst and every entry inherited that burst's timing.
2,038 loads in the first second is the cold start. The bumps at 60, 120 and 180 are the same entries expiring together, because they were all stamped at the same time.
Plot loads per second against time after a deploy. If you see bumps at multiples of your time to live, your entries are synchronised, and no amount of single flight will flatten them.