nilmdb/docs/design.md

Structure
---------
nilmdb.nilmdb is the NILM database interface.  A nilmdb.BulkData
interface stores data in flat files, and a SQL database tracks
metadata and ranges.

Access to the nilmdb must be single-threaded.  This is handled with
the nilmdb.serializer class.  In the future this could probably
be turned into a per-path serialization.

nilmdb.server is a HTTP server that provides an interface to talk,
thorugh the serialization layer, to the nilmdb object.

nilmdb.client is a HTTP client that connects to this.

Sqlite performance
------------------

Committing a transaction in the default sync mode (PRAGMA synchronous=FULL)
takes about 125msec.  sqlite3 will commit transactions at 3 times:

1. explicit con.commit()

2. between a series of DML commands and non-DML commands, e.g.
   after a series of INSERT, SELECT, but before a CREATE TABLE or
   PRAGMA.

3. at the end of an explicit transaction, e.g. "with self.con as con:"

To speed up testing, or if this transaction speed becomes an issue,
the sync=False option to NilmDB will set PRAGMA synchronous=OFF.


Inserting streams
-----------------

We need to send the contents of "data" as POST.  Do we need chunked
transfer?

- Don't know the size in advance, so we would need to use chunked if
  we send the entire thing in one request.
- But we shouldn't send one chunk per line, so we need to buffer some
  anyway; why not just make new requests?
- Consider the infinite-streaming case, we might want to send it
  immediately?  Not really -- server still should do explicit inserts
  of fixed-size chunks.
- Even chunked encoding needs the size of each chunk beforehand, so
  everything still gets buffered.  Just a tradeoff of buffer size.

Before timestamps are added:

- Raw data is about 440 kB/s    (9 channels)
- Prep data is about 12.5 kB/s  (1 phase)
- How do we know how much data to send?

    - Remember that we can only do maybe 8-50 transactions per second on
      the sqlite database.  So if one block of inserted data is one
      transaction, we'd need the raw case to be around 64kB per request,
      ideally more.
    - Maybe use a range, based on how long it's taking to read the data
        - If no more data, send it
        - If data > 1 MB, send it
    - If more than 10 seconds have elapsed, send it
    - Should those numbers come from the server?

Converting from ASCII to PyTables:

- For each row getting added, we need to set attributes on a PyTables
  Row object and call table.append().  This means that there isn't a
  particularly efficient way of converting from ascii.
- Could create a function like nilmdb.layout.Layout("foo".fillRow(asciiline)
    - But this means we're doing parsing on the serialized side
    - Let's keep parsing on the threaded server side so we can detect
      errors better, and not block the serialized nilmdb for a slow
      parsing process.
- Client sends ASCII data
- Server converts this ACSII data to a list of values
    - Maybe:

            # threaded side creates this object
            parser = nilmdb.layout.Parser("layout_name")
            # threaded side parses and fills it with data
            parser.parse(textdata)
            # serialized side pulls out rows
            for n in xrange(parser.nrows):
                parser.fill_row(rowinstance, n)
                table.append()


Inserting streams, inside nilmdb
--------------------------------

- First check that the new stream doesn't overlap.
    - Get minimum timestamp, maximum timestamp from data parser.
        - (extend parser to verify monotonicity and track extents)
    - Get all intervals for this stream in the database
    - See if new interval overlaps any existing ones
        - If so, bail
    - Question: should we cache intervals inside NilmDB?
        - Assume database is fast for now, and always rebuild fom DB.
        - Can add a caching layer later if we need to.
    - `stream_get_ranges(path)` -> return IntervalSet?

Speed
-----

- First approach was quadratic.  Adding four hours of data:

        $ time zcat /home/jim/bpnilm-data/snapshot-1-20110513-110002.raw.gz | ./nilmtool.py insert -s 20110513-110000 /bpnilm/1/raw
        real    24m31.093s
        $ time zcat /home/jim/bpnilm-data/snapshot-1-20110513-110002.raw.gz | ./nilmtool.py insert -s 20110513-120001 /bpnilm/1/raw
        real    43m44.528s
        $ time zcat /home/jim/bpnilm-data/snapshot-1-20110513-110002.raw.gz | ./nilmtool.py insert -s 20110513-130002 /bpnilm/1/raw
        real    93m29.713s
        $ time zcat /home/jim/bpnilm-data/snapshot-1-20110513-110002.raw.gz | ./nilmtool.py insert -s 20110513-140003 /bpnilm/1/raw
        real    166m53.007s

- Disabling pytables indexing didn't help:

        real    31m21.492s
        real    52m51.963s
        real    102m8.151s
        real    176m12.469s

- Server RAM usage is constant.

- Speed problems were due to IntervalSet speed, of parsing intervals
  from the database and adding the new one each time.

    - First optimization is to cache result of `nilmdb:_get_intervals`,
      which gives the best speedup.

    - Also switched to internally using bxInterval from bx-python package.
      Speed of `tests/test_interval:TestIntervalSpeed` is pretty decent
      and seems to be growing logarithmically now.  About 85μs per insertion
      for inserting 131k entries.

    - Storing the interval data in SQL might be better, with a scheme like:
      http://www.logarithmic.net/pfh/blog/01235197474

- Next slowdown target is nilmdb.layout.Parser.parse().
    - Rewrote parsers using cython and sscanf
    - Stats (rev 10831), with `_add_interval` disabled

        layout.pyx.Parser.parse:128        6303 sec, 262k calls
         layout.pyx.parse:63               13913 sec, 5.1g calls
        numpy:records.py.fromrecords:569   7410 sec, 262k calls

- Probably OK for now.

- After all updates, now takes about 8.5 minutes to insert an hour of
  data, constant after adding 171 hours (4.9 billion data points)

- Data set size: 98 gigs = 20 bytes per data point.
  6 uint16 data + 1 uint32 timestamp = 16 bytes per point
  So compression must be off -- will retry with compression forced on.

IntervalSet speed
-----------------
- Initial implementation was pretty slow, even with binary search in
  sorted list

- Replaced with bxInterval; now takes about log n time for an insertion
    - TestIntervalSpeed with range(17,18) and profiling
        - 85 μs each
        - 131072 calls to `__iadd__`
        - 131072 to bx.insert_interval
        - 131072 to bx.insert:395
        - 2355835 to bx.insert:106  (18x as many?)

- Tried blist too, worse than bxinterval.

- Might be algorithmic improvements to be made in Interval.py,
  like in `__and__`

- Replaced again with rbtree.  Seems decent.  Numbers are time per
  insert for 2**17 insertions, followed by total wall time and RAM
  usage for running "make test" with `test_rbtree` and `test_interval`
  with range(5,20):
    - old values with bxinterval:
      20.2 μS, total 20 s, 177 MB RAM
    - rbtree, plain python:
      97 μS, total 105 s, 846 MB RAM
    - rbtree converted to cython:
      26 μS, total 29 s, 320 MB RAM
    - rbtree and interval converted to cython:
      8.4 μS, total 12 s, 134 MB RAM

- Would like to move Interval itself back to Python so other
  non-cythonized code like client code can use it more easily.
  Testing speed with just `test_interval` being tested, with
  `range(5,22)`, using `/usr/bin/time -v python tests/runtests.py`,
  times recorded for 2097152:
    - 52ae397 (Interval in cython):
	  12.6133 μs each, ratio 0.866533, total 47 sec, 399 MB RAM
	- 9759dcf (Interval in python):
	  21.2937 μs each, ratio 1.462870, total 83 sec, 1107 MB RAM
  That's a huge difference!  Instead, will keep Interval and DBInterval
  cythonized inside nilmdb, and just have an additional copy in
  nilmdb.utils for clients to use.

Layouts
-------
Current/old design has specific layouts: RawData, PrepData, RawNotchedData.
Let's get rid of this entirely and switch to simpler data types that are
just collections and counts of a single type.  We'll still use strings
to describe them, with format:

    type_count

where type is "uint16", "float32", or "float64", and count is an integer.

nilmdb.layout.named() will parse these strings into the appropriate
handlers.  For compatibility:

    "RawData" == "uint16_6"
    "RawNotchedData" == "uint16_9"
    "PrepData" == "float32_8"


BulkData design
---------------

BulkData is a custom bulk data storage system that was written to
replace PyTables.  The general structure is a `data` subdirectory in
the main NilmDB directory.  Within `data`, paths are created for each
created stream.  These locations are called tables.  For example,
tables might be located at

    nilmdb/data/newton/raw/
    nilmdb/data/newton/prep/
    nilmdb/data/cottage/raw/

Each table contains:

- An unchanging `_format` file (Python pickle format) that describes
  parameters of how the data is broken up, like files per directory,
  rows per file, and the binary data format

- Hex named subdirectories `("%04x", although more than 65536 can exist)`

- Hex named files within those subdirectories, like:

        /nilmdb/data/newton/raw/000b/010a

    The data format of these files is raw binary, interpreted by the
    Python `struct` module according to the format string in the
    `_format` file.

- Same as above, with `.removed` suffix, is an optional file (Python
  pickle format) containing a list of row numbers that have been
  logically removed from the file.  If this range covers the entire
  file, the entire file will be removed.

- Note that the `bulkdata.nrows` variable is calculated once in
  `BulkData.__init__()`, and only ever incremented during use.  Thus,
  even if all data is removed, `nrows` can remain high.  However, if
  the server is restarted, the newly calculated `nrows` may be lower
  than in a previous run due to deleted data.  To be specific, this
  sequence of events:

    - insert data
    - remove all data
    - insert data

    will result in having different row numbers in the database, and
    differently numbered files on the filesystem, than the sequence:

    - insert data
    - remove all data
    - restart server
    - insert data

    This is okay!  Everything should remain consistent both in the
    `BulkData` and `NilmDB`.  Not attempting to readjust `nrows` during
    deletion makes the code quite a bit simpler.

- Similarly, data files are never truncated shorter.  Removing data
  from the end of the file will not shorten it; it will only be
  deleted when it has been fully filled and all of the data has been
  subsequently removed.


Rocket
------

Original design had the nilmdb.nilmdb thread (through bulkdata)
convert from on-disk layout to a Python list, and then the
nilmdb.server thread (from cherrypy) converts to ASCII.  For at least
the extraction side of things, it's easy to pass the bulkdata a layout
name instead, and have it convert directly from on-disk to ASCII
format, because this conversion can then be shoved into a C module.
This module, which provides a means for converting directly from
on-disk format to ASCII or Python lists, is the "rocket" interface.
Python is still used to manage the files and figure out where the
data should go; rocket just puts binary data directly in or out of
those files at specified locations.

Before rocket, testing speed with uint16_6 data, with an end-to-end
test (extracting data with nilmtool):

- insert: 65 klines/sec
- extract: 120 klines/sec

After switching to the rocket design, but using the Python version
(pyrocket):

- insert: 57 klines/sec
- extract: 120 klines/sec

After switching to a C extension module (rocket.c)

- insert: 74 klines/sec through insert.py; 99.6 klines/sec through nilmtool
- extract: 335 klines/sec

After client block updates (described below):

- insert: 180 klines/sec through nilmtool (pre-timestamped)
- extract: 390 klines/sec through nilmtool

Using "insert --timestamp" or "extract --bare" cuts the speed in half.

Blocks versus lines
-------------------

Generally want to avoid parsing the bulk of the data as lines if
possible, and transfer things in bigger blocks at once.

Current places where we use lines:

- All data returned by `client.stream_extract`, since it comes from
  `httpclient.get_gen`, which iterates over lines.  Not sure if this
  should be changed, because a `nilmtool extract` is just about the
  same speed as `curl -q .../stream/extract`!

- `client.StreamInserter.insert_iter` and
  `client.StreamInserter.insert_line`, which should probably get
  replaced with block versions.  There's no real need to keep
  updating the timestamp every time we get a new line of data.

  - Finished.  Just a single insert() that takes any length string and
    does very little processing until it's time to send it to the
	server.

Timestamps
----------

Timestamps are currently double-precision floats (64 bit).  Since the
mantissa is 53-bit, this can only represent about 15-17 significant
figures, and microsecond Unix timestamps like 1222333444.000111 are
already 16 significant figures.  Rounding is therefore an issue;
it's hard to sure that converting from ASCII, then back to ASCII,
will always give the same result.

Also, if the client provides a floating point value like 1.9999999999,
we need to be careful that we don't store it as 1.9999999999 but later
print it as 2.000000, because then round-trips change the data.

Possible solutions:

- When the client provides a floating point value to the server,
  always round to the 6th decimal digit before verifying & storing.
  Good for compatibility and simplicity.  But still might have rounding
  issues, and clients will also need to round when doing their own
  verification.  Having every piece of code need to know which digit
  to round at is not ideal.

- Always store int64 timestamps on the server, representing
  microseconds since epoch.  int64 timestamps are used in all HTTP
  parameters, in insert/extract ASCII strings, client API, commandline
  raw timestamps, etc.  Pretty big change.

  This is what we'll go with...

  - Client programs that interpret the timestamps as doubles instead
    of ints will remain accurate until 2^53 microseconds, or year
    2255.

  - On insert, maybe it's OK to send floating point microsecond values
    (1234567890123456.0), just to cope with clients that want to print
    everything as a double.  Server could try parsing as int64, and if
    that fails, parse as double and truncate to int64.  However, this
    wouldn't catch imprecise inputs like "1.23456789012e+15".  But
    maybe that can just be ignored; it's likely to cause a
    non-monotonic error at the client.

  - Timestamps like 1234567890.123456 never show up anywhere, except
    for interfacing to datetime_tz etc.  Command line "raw timestamps"
    are always printed as int64 values, and a new format
    "@1234567890123456" is added to the parser for specifying them
    exactly.

Binary interface
----------------

The ASCII interface is too slow for high-bandwidth processing, like
sinefits, prep, etc.  A binary interface was added so that you can
extract the raw binary out of the bulkdata storage.  This binary is
a little-endian format, e.g. in C a uint16_6 stream would be:

    #include <endian.h>
    #include <stdint.h>
    struct {
        int64_t timestamp_le;
        uint16_t data_le[6];
    } __attribute__((packed));

Remember to byteswap (with e.g. `letoh` in C)!

This interface is used by the new `nilmdb.client.numpyclient.NumpyClient`
class, which is a subclass of the normal `nilmcb.client.client.Client`
and has all of the same functions.  It adds three new functions:

- `stream_extract_numpy` to extract data as a Numpy array

- `stream_insert_numpy` to insert data as a Numpy array

- `stream_insert_numpy_context` is the context manager for
  incrementally inserting data

It is significantly faster!  It is about 20 times faster to decimate a
stream with `nilm-decimate` when the filter code is using the new
binary/numpy interface.


WSGI interface & chunked requests
---------------------------------

mod_wsgi requires "WSGIChunkedRequest On" to handle
"Transfer-encoding: Chunked" requests.  However, `/stream/insert`
doesn't handle this correctly right now, because:

- The `cherrpy.request.body.read()` call needs to be fixed for chunked requests

- We don't want to just buffer endlessly in the server, and it will
  require some thought on how to handle data in chunks (what to do about
  interval endpoints).

It is probably better to just keep the endpoint management on the client
side, so leave "WSGIChunkedRequest off" for now.