Let $G$ be an unweighted undirected graph with $n$ vertices and $m$ edges. Is it possible to preprocess $G$ and produce a data structure of size $m \cdot \mathrm{polylog}(n)$ so that it can answer queries of the form "distance between $u$ and $v$" in time O(n)?
The problem seems too basic to be unsolved.
This is a very interesting question. At a high level, you are asking whether one can preprocess a graph such that shortest path queries become independent of the density of the graph, without using much extra space -- interesting, but as you say, unresolved.
If you are happy with approximate distances, here is a way to get a $2$-approximation. Let $G$ be a weighted undirected graph with $n$ nodes and $m$ edges. It is shown in the following paper that for approximate distance queries, designing data structures for graphs with $m$ edges is no harder than graphs in which each node has degree bounded by $m/n$:
R. Agarwal, P. B. Godfrey, S. Har-Peled, Approximate distance queries and compact routing in sparse graphs, INFOCOM 2011
So, assume that $G$ is a $m/n$-degree bounded graph. Sample $\alpha = O(m/n)$ nodes uniform randomly; call these landmark nodes. During the preprocessing phase, store the distance from each landmark node to each other node in the graph; this requires $O(m)$ space. For each node $u$, store its closest landmark node $\ell(u)$. Also, store the graph within the data structure, say as an adjacency list.
When queried for distance between $u$ and $v$, grow balls around both nodes -- ball of node $w$ is defined as the set of nodes that are strictly closer to $w$ than to its closest landmark node, say $\ell(w)$. It can be shown that the size of each ball is $O(n^2/m)$, in expectation. Let $\Gamma(u) = B(u) \cup N(B(u))$, where $B(u)$ is the ball of node $u$ and $N(B(u))$ is the set of neighbors of nodes in $B(u)$. It can be shown that the size of $\Gamma(u)$ is $O(n)$, in expectation.
Answering the query: if $\Gamma(u) \cap \Gamma(v) \neq \emptyset$, return $\min_{x \in \Gamma(u) \cap \Gamma(v)} \{d(u, x) + d(v, x)\}$; else if $d(u, \ell(u)) \leq d(v, \ell(v))$, return $d(u, \ell(u)) + d(\ell(u), v)$; else return $d(v, \ell(v)) + d(\ell(v), u)$. It is easy to show that this is a $2$-approximation.
In terms of query time, note that growing balls takes $O(n)$ time for a $m/n$-degree bounded graph; constructing $\Gamma(u)$ and $\Gamma(v)$ given respective balls takes $O(n)$ time (since neighbors are stored within the data structure); and checking whether $\Gamma(u) \cap \Gamma(v)$ is empty or not also takes $O(n)$ time.
The above bounds are in expectation; I think it is easy to derandomize the construction. Unfortunately, this technique does not seem to allow getting approximation better than $2$. Its a very interesting question though ....
What you need is called a "distance oracle". Unfortunately, I am not very familiar with distance oracles, so I can only refer you to the seminal paper due to Thorup and Zwick:
Mikkel Thorup and Uri Zwick. Approximate distance oracles. STOC '01, 2001.
Here is an excerpt from the abstract:
Let $G = (V, E)$ be an undirected weighted graph with $|V| = n$ and $|E| = m$. Let $k$ be an integer. We show that $G = (V, E)$ can be preprocessed in $O(kmn^{1/k})$ expected time, constructing a data structure of size $O(kn^{1+1/k})$, such that any subsequent distance query can be answered, approximately, in $O(k)$ time. The approximate distance returned is of stretch at most $2k - 1$, i.e., the quotient obtained by dividing the estimated distance by the actual distance lies between 1 and $2k - 1$. [...] The space requirement of our algorithm is [...] essentially optimal.
According to their results, what you request is basically doable even for weighted graphs: choosing $k=1$ yields a distance oracle of size $O(n^2)$ obtained in expected time $O(mn)$, which can answer your shortest path queries with $1$-stretch in $O(1)$ time!
Distance oracles is a very well researched field, so you will be able to dig further I believe.
