If you've never written a compiler from scratch, I highly suggest doing so(1): there are data structures (symbol table -- a hash table, parse tree, tries), algorithms (converting an NDFSM to a DFSM(2), parsing, filling registers and many others), computer organization / architecture (outputting assembly, optimizing the code with pipelining and CPU cache in mind). Not to mention it's also a substantial project which teaches you a lot about software engineering: it's not something you can write in a single session, it's something that has to be frequently extended and tested (I can't think of a cleaner candidate for test driven development: you specify the source code you want to input and test if the generated IR/assembly code is what you're looking for). It's also really fun :-)
Not to mention custom compilers/languages are very useful in distributed systems: look at Thrift/Protocol Buffers ( languages for describing serialization of data structures for RPC). Google also has custom language which is compiled into state machines that they used to implement their Chubby service (and probably other distributed systems) -- see the "Paxos Live" paper. At an employer I've been at, which (when I was there) was deploying distributed systems with tens of thousands of nodes, we've created a custom language for describing/querying groups of machines: this language and systems built on top of it (for monitoring, configuration, provisioning and deployment), helped reduce admin:machine ratio by orders of magnitude (think a team of three handling ~10,000 machines).
(1) I suggest picking up an older (not the newest) version of the Dragon Book and targeting a x86/amd64/MIPS: while for a practical compiler I'd almost certainly suggest targeting LLVM or JVM/CLR, you'll learn more targeting bare metal.
(2) Incidentally, Lamport's landmark paper suggested treating processes in distributed systems as state machines.
I am not able to wrap my head around what was it specifically that you needed a language for - rather than, say building it on top of perl/python/lua - and whether you used the yacc/bison toolchain to build it.
The queries you'd make would be like: "Give me difference of (union of ( all machines running 64 bit Linux , all machine in cluster "match", all machines in wast coast datacenter )) machines starting "box10" and then show me the vlans these machines are on". Of course this would be in a much more terse (but very easy to learn) language. Perl/Python/C code (and through a CLI tool, shell scripts) would then do operations on returned values (i.e. parallel ssh to execute a command, parallel TCP call).
I have a Googler friend (genius coder in his own right) who sometimes wondered if he wouldn't be more productive by just devoting his workday to ensuring that Jeff Dean was properly caffeinated.
I find this kind of fascinating because Jeff Dean's academic background was in compiler optimization research. Not the obvious choice to build infrastructure for a large website. But perhaps compiler people know how every nanosecond counts, and can see the network as just another high latency part of a big computation.