Latency

A total made of parts, each with a different cause and each yielding to a different remedy or to none. Propagation is the time a signal spends travelling through fibre, which is set by distance and is the component colocation exists to remove. Processing is the time systems spend reading, validating and routing the instruction. Queueing is the time it waits behind other instructions when a system is busy, which is worst exactly when a market is moving. And the client's own connection contributes a share nobody at the venue can influence.

It runs in both directions and the return leg matters as much as the outbound one. Quotes arrive at a screen after the same delay in reverse, so a price a reader is looking at describes the market as it was rather than as it is, and an instruction sent against it is answered against the market as it has since become. That round trip is the mechanism behind a requote and behind a share of ordinary slippage, neither of which requires anyone to have behaved badly.

Published latency figures are read with care because they are averages of a distribution with a long tail. Most instructions are dealt against a book that is barely moving and complete quickly; a small minority sit far out in the tail, and the mean is dominated by the quiet majority and moves very little whatever the tail does. The tail is where the cost lands, so a mean quoted without a distribution answers a question nobody asked.