TITLE: For which metric spaces is Gromov-Hausdorff distance actually achieved?
QUESTION [10 upvotes]: Question

For which pairs $M,N$ of compact metric spaces does there exist a metric space $K$ along with isometric embeddings $i:M \to K$ and $j:N \to K$ so that the Hausdorff distance between $i(M)$ and $j(N)$ in $K$ equals the Gromov-Hausdorff distance between $M$ and $N$ exactly? 

We always have $d_\text{GH}(M,N) \leq d_\text{H}(i(M),j(N))$ by definition, so it suffices to decide if there exist any constraints intrinsic to $M$ and $N$ which force the reverse inequality to also hold for some judicious choice of $i,j,K$.
Update
Bill Johnson's nice answer appears to settle the question for compact $M$ and $N$ affirmatively: the Gromov-Hausdorff distance can always be realized by an ultraproduct construction. In light of that answer -- and also in light of the fact that I did not intend to assume compactness in the original question but stupidly wrote it down anyway -- here is a modified question:

Can the GH-distance be similarly achieved if we only assume that $M$ and $N$ are bounded, rather than compact?

In general, I am interested in the weakest hypotheses on $M$ and $N$ which are known to guarantee embeddings into a common target space $K$ so that the Gromov-Hausdorff distance between $M$ and $N$ equals the Hausdorff distance of their isometric images in $K$. It is not clear to me that a modification of Bill's argument works when we drop compactness.
Background
Given a compact metric space $M$ and two subspaces $A, B \subset M$, their Hausdorff distance -- denoted $d_\text{H}(A,B)$ -- is defined to be the smallest $\epsilon > 0$ so that $B$ is contained in $A$ thickened by $\epsilon$ and vice-versa. 
The Gromov-Hausdorff distance $d_\text{GH}(M,N)$ between two compact metric spaces $M$ and $N$ is defined to be the infimum over triples $(i,j,K)$ of $d_\text{H}(i(M),j(N))$ where $K$ is a metric space and $i,j$ are isometric embeddings of $M, N$ into $K$. The question asks when this infimum can be explicitly realized.

REPLY [2 votes]: Bounded is not sufficient to ensure that the Gromov-Hausdorff distance is actually achieved. Here is a counterexample:
Let $M$ be a metric space whose underlying set is $\mathbb{N}\times\{0,1\}$ with the following metric: $d((a,i),(b,j))=1$ if $a\neq b$ and $d((a,0),(a,1))=2^{-a}$. Let $N$ be $M$ with a single new point $x$ such that $d(x,y)=1$ for all $y\neq x$. $d_{GH}(M,N)=0$ but this cannot be realized in any common embedding because $M$ and $N$ are not isomorphic.
For bounded metric spaces we can use the formalism of continuous logic to state a few situations in which we can ensure that the distance is actually achieved (although these conditions are probably not very useful).
The first one is for any bounded metric spaces $M$ and $N$ there are elementary extensions $M^\prime \succeq M$ and $N^\prime \succeq N$ and a common embedding $i:M^\prime \rightarrow K $ and $j:N^\prime \rightarrow K$ such that $d_H(i(M^\prime),j(M^\prime))=d_{GH}(M,N)$. You can achieve this with exactly the same argument as in Bill's answer, specifically an ultrapower construction, the only difference is that for non-compact spaces the ultrapower will grow.
The second one is that if $M$ and $N$ are "resplendent" in the model theoretic sense then there will always exist $i:M\rightarrow K$ and $j:N\rightarrow K$ such that $d_H(i(M),j(N))=d_{GH}(i(M),j(N))$. Nobody has bothered to write up the details of how resplendence works in continuous logic but I'm fairly confident that the most common conditions ensuring it will work the same as they do in discrete logic, in particular if you know that $M$ and $N$ have the same density character and are either saturated or special then the Gromov-Hausdorff distance between them will be achieved.
In general though I imagine that for an arbitrary pair of bounded metric spaces this problem is very hard.