The idea is, that we have data that is obtained or calculated to give us on demand as much of it as we request. But it is not necessarily initially present. This concept is quite common in the functional world, where we in a way hide the deprecated concept of state in such structures, by the way in a way that lets use retain the benefits that led to the desire for statelessness.
Actually the concept is quite old. We have it for I/O in Unix and hence in Linux since the 1970ies. „Everything is a file“, at least as long as we constrain ourselves to a universal subset of possible file operations. It can be keyboard input, a named or anonymous pipe, an actual file, a TCP-connection, to name the most important cases. These are „lazy“ files, behave more or less like files as far as sequential reading is concerned, but not for random access reading. The I/O-concept has been done in such a way that it takes the case into account that we want to read bytes, but get only bytes. This can happen with files when we reach their end, but then we can obtain an indication that we reached the end of the file, while it is perfectly possible that we read less then we want in one access, but eventually get bytes including subsequent reads. Since the API has been done right, but by no means ideal, it generalizes well to the different cases that exist in current OS environments.
We could consider a File as an array of bytes. There is actually a way to access it in this way by memory-mapping it, but this assumes a physically present file. Now we could assume that we think of the array as a list that is optimized for sequential access and iterating, but not for random access. Both list types actually exist in languages like Java. Actually the random access structure can be made lazy as well, within certain constraints. If the source is actually sequential, we can just assume that the data is obtained up to the point where we actually read. The information about the total length of the stream may or may not be available, it is always available somehow in the case of structures that are completely available in memory. This random access on lazy collections works fine if the reason of laziness is to actually save us from doing expensive operations to obtain data that we do not actually need or to obtain them in parallel to the computation that processes the data. But we loose another potential drawback in this case. If the data is truly sequential, we can actually process data that is way beyond our memory capacity.
So the concept transfers easily from I/O-streams to lists and even arrays, most naturally to iterables that can be iterated only once. But we can easily imagine that this also applies to Strings, which can be seen a sequence of characters. If we do not constrain us to what a String is in C or Java or Ruby, but consider String to be a more abstract concept, again possibly dropping the idea of knowing the length or having a finite length. Just think of the output of the Unix command „yes“ or „cat /dev/zero“, which is infinite, in a theoretical way, but the computer won’t last forever in real life, of course. And we always interrupt the output at some time, usually be having the consumer shut down the connection.
Even numbers can be infinite. For real numbers this can happen only after the decimal point, for p-adic numbers it happens only before the decimal point, if you like to look into that. Since we rarely program with p-adic numbers this is more or less an edge case that is not part of our daily work, unless we actually do math research. But we could have integers with so many digits that we actually obtain and process them sequentially.
Reactive programming, which is promoted by lightbend in the Reactive Manifesto relies heavily on lazy structures, in this case data streams. An important concept is the so called „backpressure“, that allows the consumer to slow down the producer, if it cannot read the data fast enough.
Back to the collections, we can observe different approaches. Java 8 has introduced streams as lazy collections and we need to transform collections into streams and after the operation a stream back into a collection, at least in many real life situations. But putting all into one structure has some drawbacks as well. But looking at it from an abstract point of view this does not matter. The java8-streams to not implement a collection interface, but they are lazy collections from a more abstract point of view.
It is interesting that this allows us to relatively easily write nested loops where the depth of the nesting is a parameter that is not known at compile time. We just need a lazy collections of -tuples, where is the actual depth of the nesting and the contents are according to what the loops should iterate through. In this case we might or might not know the size of the collection, possibly not fitting into a 32-bit-integer. We might be able to produce a random member of the collection. And for sure we can iterate through it and stop the iteration wherever it is, once the desired calculation has been completed.