A team says memory beats disk by orders of magnitude, so a shared tier will cut latency; what must that claim name first?

## The claim, and the three things it hides 'Memory is orders of magnitude faster than disk' is true of the media in isolation and says almost nothing about a system until three things are named. 1. **Which medium is on the other side.** A spinning disk, a solid-state device and a page already resident in the source's own memory are three different baselines, separated by orders of magnitude from each other and not only from RAM. 2. **Whether the source is warm.** A disk engine keeps recently used pages in an in-memory area. If the pages this workload touches are already there, the comparison is memory against memory, and the medium advantage has already been collected - by the engine, without a new component. 3. **Whether a network hop is added.** A shared in-memory tier is a separate component with its own address. Reaching it costs a round trip. If the existing source is also one hop away, both end-to-end latencies are dominated by the same hop. ## The comparison once the baseline is named | baseline for one small read | rough end-to-end cost | what dominates | |---|---|---| | cold page on a spinning disk | several milliseconds | seek and rotation | | cold page on a solid-state device | hundreds of microseconds | device access | | resident page in the source, one hop away | a few hundred microseconds | the hop | | entry in a shared in-memory tier, one hop away | a few hundred microseconds | the hop | | entry held inside the calling process | under a microsecond | nothing external | The first two rows are where the slogan is true and the medium argument is strong. The middle two are the common case in a service whose source is well provisioned, and there the measured difference can be **nil**. The last row is a different arrangement altogether, and it is the only one that removes the hop rather than paying it again. ## How the gap goes to zero, or below The arithmetic is unforgiving. A memory access costs about a hundred nanoseconds; a datacentre round trip costs a few hundred microseconds - roughly three orders of magnitude more. Once both paths pay a hop, the medium difference disappears inside it. The comparison can even come out **negative**: - the new component may sit further away in the network than the source does; - a request that does not find what it wants there has still spent the hop before it continues; - a resident single-entry read at the source often costs tens of microseconds of server time - small enough that removing it changes nothing a caller can perceive. None of that says the component is pointless. It says the **latency argument** for it is unsound in this regime. There are other arguments - relieving a source that is expensive per request, or holding state that has no other home - but they are different arguments resting on different evidence, and reaching a verdict on them is a separate question from stating what the medium gap actually is. ## What to ask for instead of the ratio When the claim is made, the useful follow-ups are measurements rather than opinions: - what does the source's read latency look like now, at the median and at the high percentiles? - how much of that time is device wait, how much is server work, and how much is network? - how much of what this workload touches is already resident in the source? - where would the new component sit relative to the caller - same host, same rack, same region? If most of the time is the hop, a component one hop away cannot remove it. If most of the time is device wait on cold reads, the medium argument is real and the numbers will show it plainly. ## What varies, and must be said out loud The shape of this argument is not identical for every store or every source. Stores in this class differ in whether they are reached over a network at all; the same component held inside the calling process pays no hop, which inverts the arithmetic above. Sources differ in how much memory they were given and therefore in how often they touch a device. And some stores in this class can keep colder values on a solid-state device, which puts a device access back into a component that is otherwise all memory. Scoping the claim is not hedging - it is the content. A candidate who says 'against cold reads the gap is real, against a warm source one hop away it is close to nothing, and here is the breakdown I would ask for to tell which we have' has answered the question. One who repeats the ratio has not.