Jump to: navigation, search

Purpose: Notes for the development of QDataSet. As components are implememented and matured, they should be moved to QDataSet.

Audience: developers

See also: QDataSet developer.qdataset


  1. Introduction
  2. Model Y vs T
  3. Model Flux vs Time and Energy
  4. Rank 0 Datasets
  5. Vector Time Series
  6. Join Dimensions
  7. More Properties
  8. Operators and Scripting
    1. Rank Promotion
  9. QDataSets are immutable, kinda

1. Introduction

IDL, Matlab, Java, C and Fortran all model numbers in code. They can store groups of numbers in arrays, but they do not try to model measurements of data. When building any analysis system, the developer must first create a model for storing data. This is typically done ad-hoc and often with minimal attention, leading to problems down the road.

QDataSet attempts to introduce a language-agnostic method for modeling measurements of data. Other similar systems exist, such as NetCDF, and CDF, but they have components that tend to drag them into a particular domain. QDataSet tries to be as simple as possible, but still allowing complex operations to be done. Further, QDataSet is a simple interface, easily implemented in different environments, often with code that achieves high-performance without loosing abstraction.

This poster describes most cleanly the interface: 2011 Poster A QDataSet is basically an array with properties, and some of these properties may have values that are other QDataSets.

Note: Autoplot uses an implementation of QDataSet, which may very slightly from the specification here. For example, its UNITS property is implemented with Das2 units, where a QDataSet could call for a string.

2. Model Y vs T

A simple and oft occurring data are when we collect a scalar quantity at regular intervals. For example, we want to measure plasma density at our spacecraft each minute for a day. So in this case for the density values, we have the QDataSet we will call "density." Here are some example calls to illustrate:

density.rank()   --> 1
density.length() --> 1440
density.value(0) --> 2.3 'UNITS' ) -> 'cm**-3'

This dataset is called a "rank 1" dataset, because one index (0) is used to access the numbers. Of course we want to know when each measurement was taken, so we query for a DEPEND_0 property: 'DEPEND_0' ) -> time, a rank 1 QDataSet
time.rank()      --> 1
time.value(0)    --> 1.4832294E9   (2017-01-01T10:00Z) 'UNITS' )  -->  'seconds since 1970-01-01T00:00Z'

To encourage a standard method for concisely representing these datasets, density would be printed as:


Note, independent data has no "DEPEND_0" property. Dependent data has a "DEPEND_0" property. This rank 1 dataset occupies two physical dimensions: one for density and one for time.

3. Model Flux vs Time and Energy

Now suppose we want to represent a spectrogram of flux values occurring at grid of time and energy locations.

flux.rank()  --> 2
flux.value(0,0) --> 9.04e5 'UNITS' ) --> '1/(cm^2-s-sr)' 'DEPEND_0' ) --> time 'DEPEND_1' ) --> energy

4. Rank 0 Datasets

Often it's useful to express single measurements, and rank 0 datasets are used for this. Each dataset has a slice operation which extracts the data at the slice location. So:

density.slice(0) --> '2.3 cm**-3', a rank 0 dataset

This rank 0 dataset's value can be accessed (with value()), and it still has properties.

d1= density.slice(0) 'UNITS' ) --> 'cm**-3'

The time value is still carried along with the data: 'CONTEXT_0' ) --> t1, also a rank 0 dataset.

5. Vector Time Series

The QDataSet representing a vector time series, like a three-component B field measured every second of a day, looks like so:

bgsm.rank()   -->  2 'DEPEND_0' ) --> Time[86400]
bgsm.length() -->  86400
bgsm.length(0) --> 3
bgsm.value(0,0) --> 2.338 nT
bgsm.value(0,1) --> -7.167 nT
bgsm.value(0,2) --> -0.576 nT

But how do we know which component is which? Instead of a DEPEND_1 describing the second index, we find a property BUNDLE_1. This is a dataset that contains properties for each index. So,

bds= 'BUNDLE_1' )
bds.length() -->  3'LABEL',0)  -->  "Bx (GSM)"

This dataset has dimensions [3,0].

Note sometimes it is sufficient to have just labels for each component, and a legacy form for this data is still supported, using nominal data for the three vector labels in a DEPEND_1 dataset.

6. Join Dimensions

Occasionally instruments have multiple collection modes which change over the course of collecting data each orbit. For example, from 0:00 to 2:00 we collect at 27 energies from 10 to 20000eV each minute, and then from 2:00 to 7:00 we collect 101 energies from 500eV to 100000eV, and then from 7:00 until 9:00 we collect 27 again, and then from 9:00 to 09:30 we collect 101 energies. One way to represent this data is with a join. Each mode can be thought of as a simple rank 2 spectrogram. Another index is added, so that we have an "array of" these spectrograms in a rank 3 dataset. This is the first index of the join data set:

jds.rank()  --> 3  # there are three indexes used to get to the data.
jds.length() --> 4  # three simple spectrograms, each in the same Time,Energy space'JOIN_0')  --> 'Time'   # this just is any non-empty string. 
jds.slice(0) -->  dataset[120,27]    # 120 minutes, 27 energies
jds.slice(1) -->  dataset[300,101]   # 6 hours, 101 energies
jds.slice(2) -->  dataset[120,27]
jds.slice(3) -->  dataset[300,101]
jds.slice(0).property('DEPEND_0')  --> Time[120]
jds.slice(1).property('DEPEND_0')  --> Time[300]
jds.slice(0).property('DEPEND_1')  --> Energy[27]

A join dimension can be considered an "array of" dimension, so it could be used to indicate when measurement cadence changes as well.

7. More Properties

QDataSet is intended to represent everything fully-described science data, to a simple array of numbers, to a simple scalar number. For this reason, every property is optional. For example if the UNITS property is missing, then the data is dimensionless. If the SCALE_TYPE is missing, then it is left to the software to infer which axis type should be used.

Here are a few more useful properties:

  • TITLE, a longer length description of the data, one line that could be used for a plot title.
  • LABEL, a short description of the data, which could be used for an axis label.
  • UNITS, a representation of the physical unit, or units representing time, or a unit identifying nominal data.
  • VALID_MIN, the smallest value considered valid.
  • VALID_MAX, the largest value considered valid.
  • TYPICAL_MIN, nice setting for axis minimum.
  • TYPICAL_MAX, nice setting for axis maximum.
  • SCALE_TYPE, "log" or "linear"
  • CADENCE, a rank 0 dataset indicating the expected distance between measurements.
  • FORMAT, a string for formatting the data, such as '%9.2f' or '%d'

These are all dimension properties, meaning they describe data within one dimension.

There are structural properties as well, which are the connections between QDataSets. We've seen already:

  • DEPEND_i, declare the dependence of the ith index on another dataset.
  • BUNDLE_i, describe the data that are bundled together with the ith index.

And there are a few more:

  • DELTA_PLUS, add this to the data to indicate the one-sigma error level.
  • DELTA_MINUS, subtract this from the data to indicate the one-sigma error level.
  • BIN_PLUS, add this to get the upper limit of the 100% confidence level (e.g. time tags).
  • BIN_MINUS, subtract this to get the lower limit of the 100% confidence level (e.g. time tags).
  • BIN_MAX, the upper limit of the 100% confidence level. (Use this instead of BIN_PLUS.)
  • BIN_MIN, the limit limit of the 100% confidence level. (Use this instead of BIN_PLUS.)
  • BINS_i, these are labels for each of the indices, typically BINS_1='min,max' in ds[n,2]

Note that just because a property exists, a code might not be using the property. For example, the total function should return a QDataSet with the same units as its input, but it might not have implemented a check for VALID_MIN. The SeriesRenderer doesn't check for JOIN_0, but this would be a reasonable thing to expect. (And kudos to the fellow who goes in and implements this!)

8. Operators and Scripting

Autoplot's Jython introduces many operators for QDataSets, providing an environment similar to IDL and Matlab. Different from IDL and Matlab are the properties and metadata that make Jython scripts more concise and easier to verify. For example,

bgse= getDataSet( 'vap+cdf:' )
bz= bgse[:,2]
bmag= getDataSet( '' )
plot( bz / bmag )
angle= toDegrees( acos( bz/bmag ) )
r= where( 15 ) )

bz/bmag is a dimensionless quantity with the label "(Bz GSE)/(<|B|> (nT))", and is still has DEPEND_0 equal to time.

8.1. Rank Promotion

Note there is logic for reconciling two datasets with unequal rank. For example, suppose we want to multiply each measurement by 10:

bgse= getDataSet( 'vap+cdf:' )
bz= bgse[:,2] * 10

Here the 10 is converted to the rank 0 dataset, and then the rank 0 dataset is repeated to make a rank 1 dataset of the same geometry.

Note too that many operators will automatically convert data to QDataSets. For example, you can say:

r= where('15 degrees') ) 

to make code more explicit. Note that if angle were in radians, that it would have been converted to degrees automatically so that the two could be compared, or an error would be thrown if the units could not be converted.

9. QDataSets are immutable, kinda

Ideally QDataSets are immutable, and a dataset can be read from but not modified. However, implementations typically extend the interface so that they can be modified and are more easily used in code where the developer knows this is safe to do. For example, DDataSet is backed by a double array, and it's convenient when implementing a result to just start plugging the numbers in. Clients receive a QDataSet and generally don't know that the dataset could actually be modified.

Here's the problem, though, jython provides mutability to its datasets by making a mutable copy if the mutators are called. I need to check on this, but my guess is that it cheats and just uses the data if it is already mutable. The cost of making needless copies is memory.

Scott and folks at Iowa finally ran into the case where this was a problem, as the reference cache was allowing its data to be mutated.

Fixing the problem would involve having (1) access control switches, and (2) locking controls. For example, a dataset could be marked as read-only, so no one could further modify it. Second, it would interesting to allow for the common case where only one party will be using a dataset, and it could get a lock and modify the data, perhaps removing it from the cache...

Also it would be nice if datasets could be documented for optimizations. For example, I have a code that is checking for the valid entries on a dataset with fill. These are all valid timetags, but the performance penalty is severe. For example:

 if ( 0==DataSetAnnotations.getInstance().getAnnotation(ds,DataSetAnnotations.ANNOTATION_INVALID_COUNT) { ... optimize ...}

An associated problem with this is that many of the operators have inefficient implementations that are there to support mutations in jython. For example, Ops.findgen(200000) actually returns a FDataSet taking all the space.

Personal tools