Whether one mirror is better

The statement: the second one is faster. For you, at that moment, quite possibly. As a property of the address, no.

What genuinely differs between addresses

• Different keys, therefore separately published descriptors.
• Different introduction points, chosen independently.
• Possibly different machines behind them, possibly the same one.
• Possibly different load, since users spread unevenly across a list.

The first two are protocol level and real. The last two are the operator's arrangement and invisible from outside.

Why your experience does not generalise

A connection to an onion service goes through a path built from relays picked partly at random, and it is longer than an ordinary Tor path. Your speed is dominated by that path. Somebody else connecting to the same address at the same second gets a different path and possibly a very different experience.

So a favourite mirror is usually a favourite circuit that you have mistaken for a property of the address. Rebuild the circuit and the ranking often inverts.

The part that is stable

If the operator runs one address on smaller hardware, or takes one out of rotation, that produces a real and persistent difference. You cannot see it, and you cannot distinguish it from circuit luck without many attempts spread over time, which is a research project rather than a decision procedure.

What people get wrong

They build a ranking out of a handful of experiences. Three attempts is not a sample when the variance between paths is this large. The confidence people have in their mirror preference is not supported by the number of observations behind it.

They also assume the ranking is shared. Advice to use the second one is advice about somebody else's circuits, and it travels badly.

How many attempts would settle it

More than anybody is going to make. To separate a real difference between addresses from circuit variance you would need many attempts per address, spread over days, with fresh circuits each time, and a consistent definition of what you were measuring. That is a study.

Nobody does that, and the people most confident about which mirror is best have usually done the least of it. Confidence here is inversely related to the number of observations, which is a bad sign in any measurement.

What we will not do

Rank them. The address page lists three strings with no labels and no order beyond the order they were given to us. Calling one primary would be a claim about their state, and we do not connect to them, so we have no state to report. See what this site does not do.

The practical answer

Use whichever one you have a correct copy of. If it is slow, get a new circuit before you switch addresses, because circuit luck is the largest term and the cheapest to change. If it still fails, try another address, since they are independent and the experiment is nearly free.

• Whether a slow page means anything
• Each mirror is its own key
• The addresses

Questions people actually ask

Which awazon mirror is fastest?

Nobody can answer that for you. Most of the speed you experience comes from the path your client built, and that path is different for every person and every attempt.

Should I switch mirrors when one is slow?

Get a fresh circuit first, because that is the biggest and cheapest variable. If it still fails, trying another address is a reasonable next step.