annotate docs/TODO.txt @ 249:1da9a9ed55a3

Slightly refactored the new trackSequenceQueryNNReporter so that it is a derived class of trackAveragingReporter. This reduces code duplication
significantly. The reporter is still accessed via the nsequence QUERY directive from the command line.

* development of functionality

** exposure of all non-write functions over Web Services

At present, the radius / counting query type isn't supported over
SOAP.  Supporting it involves changing the adb__query() exported
function, so we need to be careful; or at least defining a new
exported interface, at which point perhaps it should be designed
rather than accreted... (probably this will come naturally after the
following TODO item)

** matrix of possible queries

At the moment, there are four content-based query types, each of which
does something slightly different from what you might expect from its
name.  I think that the space of possible (sensible?) queries is
larger than this -- though working out the sensible abstraction might
have to wait for more use cases -- and also that the orthogonality of
various parameters is missing.  (e.g. a silence threshold should be
applied to all queries or none, if it makes sense at all.)

Additionally, query by key (filename) might be important.

** results

Need to sort out what the results mean; is it a similarity or a
distance score, etc.  Also, is it possible to support NN queries in a
non-Euclidean space?

** SOAP / URIs

At the moment, the query and database are referred to by paths naming
files on the SOAP server's filesystem.  This makes a limited amount of
sense for the database (though exposing implementation details of
ISMS's file system is not a great idea) but makes no sense at all for
the query.  So we need to define a query data structure that can be
serialised (preferably automatically) by SOAP for use in queries.

If we ever support inserting or other write functionality over SOAP,
this will need doing for feature files (the same as queries) and for
key lists too.

** Memory management tricks

We have a friendly memory access pattern (at least on Unixoids;
Win32's API isn't a great match for mmap(), so it is significantly
slower there).  Investigate whether madvise() tricks improve
performance on any OSes.  Also, maybe investigate a specialized use of
GetViewOfFile on win32 to make it tolerable on that platform.

** LSH

Integrate the LSH indexing with the database.  Can it be done as a
separate index file, created on demand?  What are we trying to
optimize our on-disk format for, and can it be better optimized by
having multiple files?

** RDF (not necessarily related to audioDB)

Export the results of our experiments (kept in an SQL database) as
RDF, so that people can infer stuff if they know enough about our
methods.

Possibly also write an export routine for exporting an audioDB as RDF.
And laugh hollowly as XML parsers fail completely to ingest such a
monstrous file.

* architectural issues

** more safety

A couple of areas are not yet safe against runtime faults.  

*** Large databases might well end up writing off the end of the
    various tables (e.g. track, l2norm).

*** transactionality is important; the last thing that should be
    updated on insert are the free pointers (dbH->length,
    dbH->numFiles, maybe others), so that if something goes wrong in
    the meantime the database is not in an inconsistent state.

** API vs command-line

While having a command line interface is nice, having the only way to
initialize a new audioDB instance being by faking up enough of a
command line to call our wacky constructors is less nice.
Furthermore, having the "business logic" run by the constructor is
also a little bit weird.

* regression (and other) tests

** Command line interface

There is now broad coverage of the audioDB logic, with the major
exceptions of the batch insert command, and the specifying of
different keys on import.

** SOAP

The shell's support for wait() and equivalents is limited, so there
are "sleep 1"s dotted around to attempt to avoid race conditions.
Find a better way.  Similarly, using SO_REUSEADDR in bind() is a hack
that ought not to be necessary just to run the same test twice...

** Locking

The fcntl() locking should be good enough for our uses.  Investigate
whether it is in fact robust enough (including that EAGAIN workaround
for OS X; read the kernel source to find out where that's coming from
and report it if possible).

** Benchmarks

Get together a realistic set of usage cases, preferably testing each
of the query types, and benchmark them automatically.  This is
basically a prerequisite of any performance work.

* Michael's old TODO list

audioDB FIXME:

o fix segfault when query is zero-length
:-) DONE use periodic memunmap on batch insert
o allow keys to be passed as queries
:-) DONE rename 'segments' to 'tracks' in code and help files.
o test suite
o SOAP to serialize queryFile and keyList
o SOAP to serialize files on insert / batch insert ?
:-) DONE don't overwrite existing files on db create
:-) DONE implement fcntl()-based locking.
o test locking discipline (particularly over NFS between heterogenous clients)

M. Casey 13/08/07