Programming/Kdb/Labs/Big data in kdb

Example table

Let us create and populate the trades table:

trades:([]code:`symbol$();date:`date$();time:`time$();price:`float$();volume:`short$();mktflag:`symbol$();comments:());
.[`trades;();,;(
  (`CLM16;2016.04.07;10:20:00.329;38.3  ;4h;`E;"NYMEX/CME Group trade");
  (`GCM16;2016.04.07;12:00:00.055;1239.7;6h;`E;"");
  (`GCM16;2016.04.07;12:37:02.106;1240.5;1h;`E;"");
  (`CLM16;2016.04.07;13:00:00.128;38.04 ;3h;`E;"NYMEX/CME Group trade");
  (`VXK16;2016.04.07;13:22:05.617;18.85 ;5h;`E;"");
  (`GCM16;2016.04.07;14:35:01.241;1241.2;1h;`E;"");
  (`GCM16;2016.04.08;10:13:01.048;1240f ;3h;`E;"");
  (`VXK16;2016.04.08;11:34:53.417;18.53 ;1h;`E;"Transaction represents a trade in two contract months in the same class");
  (`CLM16;2016.04.08;12:00:00.227;40.61 ;3h;`E;"NYMEX/CME Group trade");
  (`VXK16;2016.04.08;12:44:00.684;18.44 ;2h;`E;"Transaction represents a trade in two contract months in the same class");
  (`VXK16;2016.04.08;12:45:33.130;18.49 ;1h;`E;"Transaction represents a trade in two contract months in the same class");
  (`CLM16;2016.04.08;15:20:02.000;40.78 ;3h;`E;"NYMEX/CME Group trade");
  (`CLM16;2016.04.11;11:00:00.105;41.43 ;2h;`E;"NYMEX/CME Group trade");
  (`VXK16;2016.04.11;14:00:00.829;18.35 ;1h;`E;"");
  (`VXK16;2016.04.11;15:14:58.775;19.05 ;2h;`E;"");
  (`GCM16;2016.04.11;16:00:00.044;1257.9;1h;`E;"");
  (`GCM16;2016.04.11;16:28:34.311;1258.7;1h;`E;""))]

Each row represents a trade in a futures contract. The value in the code column identifies that contract. The date and time columns give the dates and times of the trades as reported by the exchange. The price column contains the trade prices per contract. The volume column contains the number of contracts traded. The name of the mktflag column stands for "market flag", which distinguishes the pit trades (in which case the value in the column is `P) from the electronic trades (in which case the value is `E). In our table, which represents a hypothetical subsample from some large data set, all the trades are electronic trades. The comments column contains arbitrary comments on the data rows as strings — not as symbols. In practice, the specific comments in our example ("NYMEX/CME Group trade", "Transaction represents a trade in two contract months in the same class") would be implemented as symbol or numeric flags rather than as strings. We chose to use the string data type to demonstrate its use in a table.

Saving a table as a single binary file

We can use the following command to save this table as a binary file:

`:tradesfile set trades

The resulting file will have 965 bytes and will reside in the current directory, which you can check with

\cd

By default this will be the directory containing the q.exe executable. To save it in a different directory, we could use something like this:

(`$":C:/Users/Paul Bilokon/Documents/tradesfile") set trades

The brackets are needed as

`$":C:/Users/Paul Bilokon/Documents/tradesfile" set trades

is interpreted by q, from right to left, as

`$(":C:/Users/Paul Bilokon/Documents/tradesfile" set trades)

which isn't what we want.

In the case of success the return is the file path as a symbol:

q)(`$":C:/Users/Paul Bilokon/Documents/tradesfile") set trades
`$":C:/Users/Paul Bilokon/Documents/tradesfile"

Loading a table from a single binary file

We can then load the table from a file either using get, in which case we can specify the new table name (by assigning to a variable with that name the result of get):

trades1:get`$":C:/Users/Paul Bilokon/Documents/tradesfile"

or using load, in which ase the loaded table will be assigned to a new variable, that variable's name matching the name of the file; in this particular case, tradesfile:

load`$":C:/Users/Paul Bilokon/Documents/tradesfile"

Splayed tables

Instead of saving the table as a single binary file, we could save it splayed, i.e. so that each column is saved in a separate binary file.

Queries often refer to a small subset of such a table's columns. Storing the columns in separate files allows kdb+ to save time by loading only some of them.

To achieve this, we would normally call set with a directory path, rather than a file path, as its first argument.

However, the following will fail with `type:

(`$":C:/Users/Paul Bilokon/Documents/tradesdir/")set trades

Unenumerated columns

set will disallow us to splay the table trades since it contains unenumerated columns of type symbol. (The same error would be thrown if the table were keyed; we would need to remove the key with

()xkey trades

before saving the table.)

We can remove those columns and save the table in one line with

(`$":C:/Users/Paul Bilokon/Documents/tradesdir/")set `code`mktflag _ trades

Notice the traling / in the directory path.

As a result of running this, the directory tradesdir will be created under C:/Users/Paul Bilokon/Documents. It will contain the files with names

• .d,
• comments,
• comments#,
• date,
• price,
• time, and
• volume.

Enumerating the columns

We couldn't save the original trades table because the columns of type symbol, namely code and mktflag, weren't enumerated. We shall now show how to enumerate them. Before proceeding, delete the directory tradesdir with all its contents, in case you created that directory by following the steps outlined above, so that we start with a clean slate. To enumerate the symbol columns in trades, we use .Q.en:

tradesenum:.Q.en[`$":C:/Users/Paul Bilokon/Documents/tradesdir/";trades]

This will create the directory tradesdir with a single binary file in it, named sym. That file enumerates the values found in the columns of type symbol, namely `CLM16, `GCM16, `VXK16, and `E.

Once this is done, we run

(`$":C:/Users/Paul Bilokon/Documents/tradesdir/")set tradesenum

The directory now contains the files with names

• .d,
• code,
• comments,
• comments#,
• date,
• mktflag,
• price,
• sym,
• time, and
• volume.

Notice that this time, with the symbol columns enumerated, we were able to save the columns of type symbol successfully.

Note that, unlike the above command,

(`$":C:/Users/Paul Bilokon/Documents/tradesdir/")set trades

will fail with 'type, as before.

Splayed tables on disc

Let us now take a look at the individual files in the directory tradesdir. The files date, price, time, and volume contain the serializations of the columns with the corresponding names.

Notice that you can serialize the columns individually. The command

(`$":C:/Users/Paul Bilokon/Documents/myprice")set trades[`price]

will create a file myprice, whose contents are identical to those of tradesdir/price.

Recall that trades[`price] is a list of type float:

q)trades[`price]
38.3 1239.7 1240.5 38.04 18.85 1241.2 1240 18.53 40.61 18.44 18.49 40.78 41.43 18.35 19.05 1257.9 1258.7
q)type trades[`price]
9h

When we run

(`$":C:/Users/Paul Bilokon/Documents/mycols")set cols trades

we obtain a file, mycols, whose contents match those of tradesdir/.d. This file specifies the order of the columns in the splayed table. (How else would we know it?)

Thus .d is the serialization of the list of column names (as symbols) in the table.

However, when we run

(`$":C:/Users/Paul Bilokon/Documents/mycode")set trades[`code]

we obtain the file mycode, whose contents differ from those of tradesdir/code. This is because all the symbols in the enumerated table are stored in tradesdir/sym, whereas tradesdir/code contains indices of the symbols within tradesdir/sym.

We will reproduce tradesdir/code by running

(`$":C:/Users/Paul Bilokon/Documents/mycode")set tradesenum[`code]

Enumerated versus unenumerated columns

This underlines the difference between trades and tradesenum: .Q.en returns a copy of the table with the symbol columns enumerated.

We can distinguish between the columns that have been enumerated and those that haven't as follows:

q)type trades[`code]
11h
q)type tradesenum[`code]
21h

although the results of meta trades and meta tradesenum look identical.

Similarly to tradesdir/code, the file tradesdir/mktflag contains serialized enumerated strings containing indices of the actual strings within tradesdir/sym.

Each distinct text string, corresponding to each distinct symbol value, is stored in tradesdir/sym once. Using the terminology from Java, symbols are interned.

Columns of strings

The column comments, whose type is string (rather than symbol) was serialized by set to two files:

• tradesdir/comments and
• tradesdir/comments#.

The latter, comments#, contains the actual text found in the strings.

Unlike the serializations of the enumerated symbol columns, where distinct values are stored only once, comments# contains repeats.

If you open this file in a binary file viewer, you will see that "NYMEX/CME Group trade", for example, occurs multiple times.

Using the terminology from Java, serialized strings are not interned, whereas serialized symbols are. The file tradesdir/comments indexes into

tradesdir/comments#

Appending to a splayed table

Now suppose that we have

moretrades:([]code:`symbol$();date:`date$();time:`time$();price:`float$();volume:`short$();mktflag:`symbol$();comments:());
.[`moretrades;();,;(
  (`GCM16;2016.04.11;16:29:05.890;1257.8;2h;`E;"");
  (`VXK16;2016.04.11;16:29:38.575;19.49 ;1h;`E;"");
  (`VXK16;2016.04.11;16:29:38.575;19.49 ;1h;`E;"");
  (`CLM16;2016.04.11;16:29:45.105;41.45 ;2h;`E;"NYMEX/CME Group trade");
  (`GCM16;2016.04.11;16:29:50.898;1257.8;1h;`E;"");
  (`GCM16;2016.04.12;10:01:05.890;1257.8;2h;`E;"");
  (`GCM16;2016.04.12;10:02:05.287;1257.8;2h;`E;"");
  (`VXK16;2016.04.12;10:03:38.123;19.49 ;1h;`E;"");
  (`VXK16;2016.04.12;10:03:38.123;19.49 ;1h;`E;"");
  (`CLM16;2016.04.12;10:18:45.230;41.45 ;2h;`E;"NYMEX/CME Group trade");
  (`GCM16;2016.04.12;10:18:50.382;1257.8;1h;`E;""))];

How can we append them to our splayed table?

If we were to do

(`$":C:/Users/Paul Bilokon/Documents/tradesdir/")set .Q.en[`$":C:/Users/Paul Bilokon/Documents/tradesdir/";moretrades]

the table would be overwritten. However, we don't want to overwrite it; we want to append to it.

What we want is

upsert[`$":C:/Users/Paul Bilokon/Documents/tradesdir/"].Q.en[`$":C:/Users/Paul Bilokon/Documents/tradesdir/";moretrades]

Sorting a splayed table

Yes, upsert works with tables on disc, not just in memory. upsert is not alone in this. For example, we can sort the splayed table on disc:

`code xasc`$":C:/Users/Paul Bilokon/Documents/tradesdir/"

We confirm that the sorted attribute is now set on the column code:

q)meta get`$":C:/Users/Paul Bilokon/Documents/tradesdir/"
c       | t f a
--------| -----
code    | s   s
date    | d
time    | t
price   | f
volume  | h
mktflag | s
comments| C

Applying attributes to a splayed table

xasc is not alone in being able to work with tables on disc. Here is how we can set the grouped (`g#) attribute on our table:

q)@[`$":C:/Users/Paul Bilokon/Documents/tradesdir/";`mktflag;`g#]
`:tradesdir/
q)meta get`$":C:/Users/Paul Bilokon/Documents/tradesdir/"
c       | t f a
--------| -----
code    | s   s
date    | d
time    | t
price   | f
volume  | h
mktflag | s   g
comments| C

Compression

Let us recreate the on-disc version of the table

trades:([]code:`symbol$();date:`date$();time:`time$();price:`float$();volume:`short$();mktflag:`symbol$();comments:());
.[`trades;();,;(
  (`CLM16;2016.04.07;10:20:00.329;38.3  ;4h;`E;"NYMEX/CME Group trade");
  (`GCM16;2016.04.07;12:00:00.055;1239.7;6h;`E;"");
  (`GCM16;2016.04.07;12:37:02.106;1240.5;1h;`E;"");
  (`CLM16;2016.04.07;13:00:00.128;38.04 ;3h;`E;"NYMEX/CME Group trade");
  (`VXK16;2016.04.07;13:22:05.617;18.85 ;5h;`E;"");
  (`GCM16;2016.04.07;14:35:01.241;1241.2;1h;`E;"");
  (`GCM16;2016.04.08;10:13:01.048;1240f ;3h;`E;"");
  (`VXK16;2016.04.08;11:34:53.417;18.53 ;1h;`E;"Transaction represents a trade in two contract months in the same class");
  (`CLM16;2016.04.08;12:00:00.227;40.61 ;3h;`E;"NYMEX/CME Group trade");
  (`VXK16;2016.04.08;12:44:00.684;18.44 ;2h;`E;"Transaction represents a trade in two contract months in the same class");
  (`VXK16;2016.04.08;12:45:33.130;18.49 ;1h;`E;"Transaction represents a trade in two contract months in the same class");
  (`CLM16;2016.04.08;15:20:02.000;40.78 ;3h;`E;"NYMEX/CME Group trade");
  (`CLM16;2016.04.11;11:00:00.105;41.43 ;2h;`E;"NYMEX/CME Group trade");
  (`VXK16;2016.04.11;14:00:00.829;18.35 ;1h;`E;"");
  (`VXK16;2016.04.11;15:14:58.775;19.05 ;2h;`E;"");
  (`GCM16;2016.04.11;16:00:00.044;1257.9;1h;`E;"");
  (`GCM16;2016.04.11;16:28:34.311;1258.7;1h;`E;""))]

but this time use compression:

(`$":C:/Users/Paul Bilokon/Documents/tradesdir/";17;2;9)set .Q.en[`$":C:/Users/Paul Bilokon/Documents/tradesdir/";trades]

Without compression the table occupied 17,595 bytes; with compression — only 1,182 bytes.

To read up on file compression in kdb+/q, see https://code.kx.com/q/kb/file-compression/.

Querying splayed tables on disc

Some (but not all) q-sql queries work with tables on disc, for example:

q)select from`$":C:/Users/Paul Bilokon/Documents/tradesdir/"where price>100f
code  date       time         price  volume mktflag comments
------------------------------------------------------------
GCM16 2016.04.07 12:00:00.055 1239.7 6      E       ""
GCM16 2016.04.07 12:37:02.106 1240.5 1      E       ""
GCM16 2016.04.07 14:35:01.241 1241.2 1      E       ""
GCM16 2016.04.08 10:13:01.048 1240   3      E       ""
GCM16 2016.04.11 16:00:00.044 1257.9 1      E       ""
GCM16 2016.04.11 16:28:34.311 1258.7 1      E       ""

However, something like

delete price from`$":tradesdir/"

wouldn't work; we would need to manipulate the table in-memory and then save down the modified table to disc, overwriting the original.

Partitioned tables

In addition to splaying a table, we may partition it: for each distinct value v of a column that defines the partition, a distinct directory will be created, where the binary files with serialized columns will be stored, but including only those rows with value v in the partition column.

The standard .Q.dpft function performs the partitioning.

The name of that function stands for its four parameters:

• directory;
• partition;
• sort field;
• table name.

To store the trades table, which is, as before,

trades:([]code:`symbol$();date:`date$();time:`time$();price:`float$();volume:`short$();mktflag:`symbol$();comments:());
.[`trades;();,;(
  (`CLM16;2016.04.07;10:20:00.329;38.3  ;4h;`E;"NYMEX/CME Group trade");
  (`GCM16;2016.04.07;12:00:00.055;1239.7;6h;`E;"");
  (`GCM16;2016.04.07;12:37:02.106;1240.5;1h;`E;"");
  (`CLM16;2016.04.07;13:00:00.128;38.04 ;3h;`E;"NYMEX/CME Group trade");
  (`VXK16;2016.04.07;13:22:05.617;18.85 ;5h;`E;"");
  (`GCM16;2016.04.07;14:35:01.241;1241.2;1h;`E;"");
  (`GCM16;2016.04.08;10:13:01.048;1240f ;3h;`E;"");
  (`VXK16;2016.04.08;11:34:53.417;18.53 ;1h;`E;"Transaction represents a trade in two contract months in the same class");
  (`CLM16;2016.04.08;12:00:00.227;40.61 ;3h;`E;"NYMEX/CME Group trade");
  (`VXK16;2016.04.08;12:44:00.684;18.44 ;2h;`E;"Transaction represents a trade in two contract months in the same class");
  (`VXK16;2016.04.08;12:45:33.130;18.49 ;1h;`E;"Transaction represents a trade in two contract months in the same class");
  (`CLM16;2016.04.08;15:20:02.000;40.78 ;3h;`E;"NYMEX/CME Group trade");
  (`CLM16;2016.04.11;11:00:00.105;41.43 ;2h;`E;"NYMEX/CME Group trade");
  (`VXK16;2016.04.11;14:00:00.829;18.35 ;1h;`E;"");
  (`VXK16;2016.04.11;15:14:58.775;19.05 ;2h;`E;"");
  (`GCM16;2016.04.11;16:00:00.044;1257.9;1h;`E;"");
  (`GCM16;2016.04.11;16:28:34.311;1258.7;1h;`E;""))]

we can execute the following code:

{
  `trd set select from trades where date=x;
  .Q.dpft[`$":C:/Users/Paul Bilokon/Documents/mydb";x;`code;`trd];
  delete trd from`.;}each distinct trades[`date]

We will thus call .Q.dpft once for each of the distinct values in the date column, resulting in the following calls:

.Q.dpft[`$":C:/Users/Paul Bilokon/Documents/mydb"; 2016.04.07; `code; `trd1];
.Q.dpft[`$":C:/Users/Paul Bilokon/Documents/mydb"; 2016.04.08; `code; `trd2];
.Q.dpft[`$":C:/Users/Paul Bilokon/Documents/mydb"; 2016.04.11; `code; `trd3];

Here `trd1`trd2`trd3 are, respectively, the names of the three tables obtained by selecting from trades the rows where date has one of the values, respectively, 2016.04.07, 2016.04.08, 2016.04.11:

select from trades where date=2016.04.07;
select from trades where date=2016.04.08;
select from trades where date=2016.04.11;

We could not implement our calls as follows

.Q.dpft[`$":C:/Users/Paul Bilokon/Documents/mydb"; 2016.04.07; `code; select from trades where date=2016.04.07];
.Q.dpft[`$":C:/Users/Paul Bilokon/Documents/mydb"; 2016.04.08; `code; select from trades where date=2016.04.08];
.Q.dpft[`$":C:/Users/Paul Bilokon/Documents/mydb"; 2016.04.11; `code; select from trades where date=2016.04.11];

because of this limitation of .Q.dpft: it expects, as its fourth argument, the table name, not the actual table.

The table is always saved to disc in a directory with that name. This is why we created the temporary table trd, with the date we will write to a partition in the global namespace (`.) using `trd set... rather than, for example, trd:. The latter will create the variable trd in the local scope of our anonymous function defined in the {...} block. We wouldn't be able to reference that variable as `trd.

At the end of our local function we delete the temporary table from the global namespace:

delete trd from `.;

The following directory structure will be obtained by evaluating our code. Under C:/Users/Paul Bilokon/Documents/mydb:

• 2016.04.07
  • trd
    • .d
    • code
    • comments
    • comments#
    • date
    • mktflag
    • price
    • time
    • volume
• 2016.04.08
  • trd
    • .d
    • ...
    • volume
• 2016.04.11
  • trd
    • .d
    • ...
    • volume
• sym

In the directory that we have specified to .Q.dpft, we have one subdirectory per each partition (2016.04.07, 2016.04.08, and 2016.04.11). Inside those directories we have a subdirectory whose name is that of the table that we provided to .Q.dpft as an argument: trd. Inside trd, we have the same files as we encountered when splaying the table, with the exception of sym, which now resides two levels above. As before, the file sym enumerates the columns of type symbol

Notice that we didn't have to call .Q.en explicitly this time: .Q.dpft did this for us.

Adding another table

We can add another table to the mydb directory.

Suppose that we have the following (again, highly contrived) table of quotes:

quotes:([]code:`symbol$();date:`date$();time:`time$();bidprice:`float$();bidsize:`short$();askprice:`float$();asksize:`short$();mktflag:`symbol$());
.[`quotes;();,;(
  (`CLM16;2016.04.07;10:15:00.010;38.34 ;86h;38.35 ;3h ;`E);
  (`GCM16;2016.04.07;11:02:16.663;1241.4;22h;1241.5;1h ;`E);
  (`CLM16;2016.04.07;12:05:00.303;38.12 ;7h ;38.13 ;13h;`E);
  (`CLM16;2016.04.07;12:22:00.486;38.11 ;16h;38.12 ;8h ;`E);
  (`GCM16;2016.04.07;13:00:00.205;1238.6;8h ;1238.7;7h ;`E);
  (`CLM16;2016.04.07;15:00:00.051;38.52 ;9h ;38.53 ;18h;`E);
  (`GCM16;2016.04.07;15:20:02.224;1240.9;6h ;1241f ;1h ;`E);
  (`CLM16;2016.04.08;10:53:00.002;40.83 ;6h ;40.84 ;66h;`E);
  (`CLM16;2016.04.08;13:56:30.070;40.54 ;38h;40.56 ;58h;`E);
  (`CLM16;2016.04.08;15:20:02.000;40.77 ;26h;40.79 ;44h;`E);
  (`CLM16;2016.04.08;15:21:43.786;40.76 ;3h ;40.77 ;28h;`E))]

We run the following code...

{
  `qt set select from quotes where date=x;
  .Q.dpft[`$":C:/Users/Paul Bilokon/Documents/mydb";x;`code;`qt];
  delete qt from `.;} each distinct quotes[`date]

...and obtain the following directory structure under C:/Users/Paul Bilokon/Documents/mydb:

• 2016.04.07
  • trd
    • .d
    • code
    • comments
    • comments#
    • date
    • mktflag
    • price
    • time
    • volume
  • qt
    • .d
    • askprice
    • asksize
    • bidprice
    • bidsize
    • code
    • date
    • mktflag
    • time
• 2016.04.08
  • trd
    • .d
    • ...
    • volume
  • qt
    • .d
    • ...
    • time
• 2016.04.11
  • trd
    • .d
    • ...
    • volume
  • qt
    • .d
    • ...
    • time
• sym

Loading partitioned tables

To load partitioned tables into kdb+, you can pass the name of the directory that contains them as a command line parameter to the q executable (we killed the already opened q instance first before restarting it in this manner):

C:\q\w64>q "C:\Users\Paul Bilokon\Documents\mydb"

Caveat

However, we are in for a surprise:

q)\v
`date`sym`trd

Where's qt? What's the matter?

It turns out that, while trd has the partitions 2016.04.07, 2016.04.08, and 2016.04.11, qt had only 2016.04.07 and 2016.04.08.

Each table in mydb must have exactly the same partitions.

We remedy things by recreating quotes and then storing quotes as qt using

{
  `qt set select from quotes where date=x;
  .Q.dpft[`$":C:/Users/Paul Bilokon/Documents/mydb";x;`code;`qt];
  delete qt from`.;} each 2016.04.07 2016.04.08 2016.04.11

On quitting kdb+ and then restarting it with

C:\q\w64>q "C:\Users\Paul Bilokon\Documents\mydb"

we see that now both trd and qt are present:

q)\v
`date`qt`sym`trd

We can examine the tables by running

select from trd

and

select from qt

The importance of current directory

Check (by running \cd) that the current directory is

"C:\\Users\\Paul Bilokon\\Documents\\mydb"

Notice that if we change it, say, with

\cd C:/q/w64

our partitioned tables stop "working":

q)select from trd
'./2016.04.07/trd/code: The system cannot find the path specified.

Changing the current directory in the presence of partitioned tables wasn't a good move. We fix things by changing it back:

\cd C:/Users/Paul Bilokon/Documents/mydb

Notice that, two other variables are present in the restarted kdb+ instance: date containing the list of the partitions:

2016.04.07 2016.04.08 2016.04.11

and sym, containing the list of enumerated symbols:

`CLM16`GCM16`VXK16`E

Appending rows to partitioned tables

Now let us suppose we want to append the following rows:

morequotes:([]code:`symbol$();date:`date$();time:`time$();bidprice:`float$();bidsize:`short$();askprice:`float$();asksize:`short$();mktflag:`symbol$());
.[`morequotes;();,;(
  (`GCM16;2016.04.07;15:23:21.147;1241.2;4h ;1241.3;1h;`E);
  (`GCM16;2016.04.07;15:33:04.535;1241.7;17h;1241.6;2h;`E))]

Naïvely, we attempt to run

qt:morequotes;
.Q.dpft[`$":C:/Users/Paul Bilokon/Documents/mydb";2016.04.07;`code;`qt];
delete qt from`.;

On restarting the kdb+ instance (we have to do this because we replaced qt with our temporary table; we wrote to qt image on disc but the in-memory representation is not updated) with

C:\q\w64>q "C:\Users\Paul Bilokon\Documents\mydb"

we find that we did not append to the 2016.04.07 partition of qt but replaced it:

select from qt

date       code  time         bidprice bidsize askprice asksize mktflag
-----------------------------------------------------------------------
2016.04.07 GCM16 15:23:21.147 1241.2   4       1241.3   1       E
2016.04.07 GCM16 15:33:04.535 1241.7   17      1241.6   2       E
2016.04.08 CLM16 10:53:00.002 40.83    6       40.84    66      E
2016.04.08 CLM16 13:56:30.070 40.54    38      40.56    58      E
2016.04.08 CLM16 15:20:02.000 40.77    26      40.79    44      E
2016.04.08 CLM16 15:21:43.786 40.76    3       40.77    28      E

This exposes another limitation of .Q.dpft: since it does not allow us to append to a partition but overwrites it, it requires the entire partition to be in memory for it to be written to a partitioned table. For some very large partitions this is infeasible: they simply won't fit in memory.

To append to a partition on disc, we need a modified version of .Q.dpft.

Let us examine the code of .Q.dpft. We see that it is written in k:

q).Q.dpft
k){[d;p;f;t;s]i:<t f;r:+enxs[$;d;;s]`. . `\:t;{[d;t;i;u;x]@[d;x;:;u t[x]i]}[d:par[d;p;t];r;i;]'[(::;`p#)f=!r;!r];@[d;`.d;:;f,r@&~f=r:!r];t}[;;;;`sym]

The part of the code that we are interested in is

@[d;x;:;u t[x]i]

here we replace : with , and obtain

q).Q.dpft
k)dpftappend:{[d;p;f;t;s]i:<t f;r:+.Q.enxs[$;d;;s]`. . `\:t;{[d;t;i;u;x]@[d;x;,;u t[x]i]}[d:.Q.par[d;p;t];r;i;]'[(::;`p#)f=!r;!r];@[d;`.d;:;f,r@&~f=r:!r];t}[;;;;`sym]

We can test it by recreating the quotes table, writing it to disc with

{
  `qt set select from quotes where date=x;
  .Q.dpft[`$":C:/Users/Paul Bilokon/Documents/mydb";x;`code;`qt];
  delete qt from`.;} each 2016.04.07 2016.04.08 2016.04.11

again, then appending morequotes with

qt:morequotes;
dpftappend[`$":C:/Users/Paul Bilokon/Documents/mydb";2016.04.07;`qt];
delete qt from`.;

(We remark that this modification of .Q.dpft was suggested by Pawan Singh — see https://groups.google.com/forum/\#!topic/personal-kdbplus/51rxAntpcOc.)

This time we have achieved our goal of appending to a partition in qt: after restarting again with

C:\q\w64>q "C:\Users\Paul Bilokon\Documents\mydb"

we check that we have indeed appended to a partition on disc:

select from qt

date       code  time         bidprice bidsize askprice asksize mktflag
-----------------------------------------------------------------------
2016.04.07 CLM16 10:15:00.010 38.34    86      38.35    3       E
2016.04.07 CLM16 12:05:00.303 38.12    7       38.13    13      E
2016.04.07 CLM16 12:22:00.486 38.11    16      38.12    8       E
2016.04.07 CLM16 15:00:00.051 38.52    9       38.53    18      E
2016.04.07 GCM16 11:02:16.663 1241.4   22      1241.5   1       E
2016.04.07 GCM16 13:00:00.205 1238.6   8       1238.7   7       E
2016.04.07 GCM16 15:20:02.224 1240.9   6       1241     1       E
2016.04.07 GCM16 15:23:21.147 1241.2   4       1241.3   1       E
2016.04.07 GCM16 15:33:04.535 1241.7   17      1241.6   2       E
2016.04.08 CLM16 10:53:00.002 40.83    6       40.84    66      E
2016.04.08 CLM16 13:56:30.070 40.54    38      40.56    58      E
2016.04.08 CLM16 15:20:02.000 40.77    26      40.79    44      E
2016.04.08 CLM16 15:21:43.786 40.76    3       40.77    28      E