Building a mini-LSST Broker for Data Management and Discovery

========

Version 0.4

By AA Miller 2017 Apr 10

Edited G. Narayan 2017 Apr 26

Edited AA Miller 2018 Feb 5

Edited AA Miller 2018 Mar 26

This notebook was developed within the Google framework. The original notebook can be run in a web browser, and is available . It has been recreated below, though we recommend you run the web-based version.

LSST will produce an unprecedented volume of astronomical time-domain information. $>37$ trillion individual photometric measurements will be recorded. While the vast, vast majority of these measurements will simply confirm the status quo, some will represent rarities that have never been seen before (e.g., LSST may be the first telescope to discover an optical orphan afterglow from a gamma-ray burst), which the community will need to know about in ~real time.

Storing, filtering, and serving this data is going to be a huge nightmare challenge. Brokers (which is an overloaded term meaning too many things), such as the ANTARES project at NOAO, are one proposed solution to this challenge. In this exercise you will build a miniature version of a broker.

Many of the difficult, and essential, steps necessary for LSST Brokers will be skipped here as they are too time consuming or beyond the scope of what we have previously covered. We will point out these challenges are we come across them.

Note

Especially for Problem 2 make an effort to include docstrings. The minimum level of docstrings should include: a short description, input parameters, object attributes (if writing a class object), and a description of what the function returns (if anything).

Warning

The next several cells contain commands that should only be run if the notebook is being run via google collaboratory.

End of Warning

Later in the notebook we will attempt to speed up our processing via a few parallel operations, the next few slides set that up.

Problem 1) Light Curve Data

We begin by ignoring the streaming aspect of the problem and instead we will work with full light curves. One way of thinking about this is that we're working in a mode where LSST has been imaging this field at least a few times, so we have some historical data on it, which we've already associated with an alert from LSST.

The collection of light curves is real multiband data, and like LSST, it features objects of different types covering a large range in brightness, and observations in multiple filters taken at different cadences.

As a first step you will create a Python class to read in the data and store light curves as python objects. The data are stored in flat text files with the following format:

and names FAKEXXX.dat where the XXX is a running index from 001 to 100.

Problem 1a

Read in the data for the first light curve file and plot the $g'$ light curve for that source.

Hint - Use pandas.

As we have many light curve files (in principle as many as 37 billion...), we will define a Source() class to ease our handling of the data.

Problem 1b

Complete the plotting method in the Source class defined below.

Hint - a for loop can be helpful for dealing with the different filters

Problem 1c

Confirm the corrections made in 1b by plotting the multiband light curve for the source FAKE010.

One thing that we brushed over previously is that the brightness measurements have units of flux, rather than the traditional use of magnitudes. The reason for this is that LSST will measure flux variations via image differencing, which will for some sources in some filters result in a measurement of negative flux. (You may have already noticed this in 1a.) Statistically there is nothing wrong with such a measurement, but it is impossible to convert a negative flux into a magnitude. Thus we will use flux measurements throughout this exercise.

Using flux measurements will allow us to make unbiased measurements of the statistical distributions of the variations of the sources we care about.

Problem 1d

What is FAKE010 the source that is plotted above?

Hint - ask Griffin or Rob...

Solution 1d

write your solution here

Problem 1e

To get a better sense of the data, plot the multiband light curves for sources FAKE060 and FAKE095.

Problem 2) Data Preparation

While we could create a database table that includes every single photometric measurement made by LSST, this ~37 trillion row db would be enormous without providing a lot of added value beyond the raw flux measurements [while this table is necessary, alternative tables may provide more useful information]. Furthermore, extracting individual light curves from such a database will be slow. Instead, we are going to develop summary statistics for every source which will make it easier to select individual sources and develop classifiers to identify objects of interest.

Below we will create a new class, Variable, which will inherit the properties of the Source class. The Variable class includes several additional methods so we can (eventually) store summary statistics in a database table. In the interest of time, we limit the summary statistics to a relatively small list all of which have been shown to be useful for classification (see Richards et al. 2011 for further details). The statistics that we include (for now) are:

1. Std -- the standard deviation of the flux measurements
2. Amp -- the amplitude of flux deviations
3. MAD -- the median absolute deviation of the flux measurements
4. beyond1std -- the fraction of flux measurements beyond 1 standard deviation
5. skew -- the skewness of a data set
6. the mean $g' - r'$, $r' - i'$, and $i' - z'$ color

Note - while the magnitude system used in astronomy is stupid, and as a reminder:

where $C$ is a constant. Comparisons of $\Delta \mathrm{flux}$ for different stars encodes both the brightness variations and the distance to the star in question. $\Delta \mathrm{mag}$, on the other hand, normalizes the brightness variations to provide a more apples to apples comparison (however the use of a logarithmic basis results in a biased apples to apples comparison).

All of this is a long way of saying - when you calculate the Std, Amp, and MAD below you want to normalize these measurements by dividing each by the weighted mean of the flux in each filter.

Problem 2a

Create the Variable class, which should inherit the properties of the Source class. Include methods within the Variable class to calculate the 6 summary statistics listed above. Be sure to point to Source explicitly, and do not use the super() method that Jeff discussed previously (this matters later on).

Hint - use python lambda functions to define the calculation, then loop over the filters. 1 example lambda function is given to calculate the weighted mean flux.

Hint 2 - use scipy.stats.skew to calculate the skew.

food for thought - if a source is observed in different filters but the observations are not simultaneous (or quasi-simultaneous), what is the meaning of a "mean color"?

Solution to food for thought - in this case we simply want you to take the mean flux in each filter and create a statistic that is $-2.5 \log \frac{\langle f_X \rangle}{\langle f_{Y} \rangle}$, where ${\langle f_{Y} \rangle}$ is the mean flux in band $Y$, while $\langle f_X \rangle$ is the mean flux in band $X$, which can be $g', r', i', z'$. Note that our use of image-difference flux measurements, which can be negative, means you'll need to add some form a case excpetion if $\langle f_X \rangle$ or $\langle f_Y \rangle$ is negative. In these cases set the color to -999.

Problem 2b

Confirm your solution to 2a by measuring the mean colors of source FAKE010. Does your measurement make sense given the plot you made in 1c?

Problem 3) Store the sources in a database

Building (and managing) a database from scratch is a challenging task. For (very) small projects one solution to this problem is to use SQLite, which is a self-contained, publicly available SQL engine. One of the primary advantages of SQLite is that no server setup is required, unlike other popular tools such as postgres and MySQL. In fact, SQLite is already integrated with python so everything we want to do (create database, add tables, load data, write queries, etc.) can be done within Python.

Without diving too deep into the details, here are situations where SQLite has advantages and disadvantages according to their own documentation:

Advantages

1. Situations where expert human support is not needed
2. For basic data analysis (SQLite is easy to install and manage for new projects)
3. Education and training

Disadvantages

1. Client/Server applications (SQLite does not behave well if multiple systems need to access db at the same time)
2. Very large data sets (SQLite stores entire db in a single disk file, other solutions can store data across multiple files/volumes)
3. High concurrency (Only 1 writer allowed at a time for SQLite)

From the (limited) lists above, you can see that while SQLite is perfect for our application right now, if you were building an actual ANTARES-like system a more sophisticated database solution would be required.

Problem 3a

Import sqlite3 into the notebook.

Hint - if this doesn't work, you may need to run

conda install sqlite3

within your IDEAS conda environment.

Following the sqlite3 import, we must first connect to the database. If we attempt a connection to a database that does not exist, then a new database is automatically created. Here we will create a new database file, called MiniBroker.db.

We now have a database connection object, conn. To interact with the database (create tables, load data, write queries) we need a cursor object.

Now that we have a cursor object, we can populate the database. As an example we will start by creating a table to hold all the raw photometry (though ultimately we will not use this table for analysis).

Note - there are many cursor methods capable of interacting with the database. The most common, execute, takes a single SQL command as its argument and executes that command. Other useful methods include executemany, which is useful for inserting data into the database, and executescript, which take an SQL script as its argument and executes the script.

In many cases, as below, it will be useful to use triple quotes in order to improve the legibility of your code.

Let's unpack everything that happened in these two commands. First - if the table rawPhot already exists, we drop it to start over from scratch. (this is useful here, but should not be adopted as general practice)

Second - we create the new table rawPhot, which has 6 columns: id - a running index for every row in the table, objId - an ID to identify which source the row belongs to, t - the time of observation in MJD, pb - the passband of the observation, flux the observation flux, and dflux the uncertainty on the flux measurement. In addition to naming the columns, we also must declare their type. We have declared id as the primary key, which means this value will automatically be assigned and incremented for all data inserted into the database. We have also declared pb as a variable character of length 1, which is more useful and restrictive than simply declaring pb as text, which allows any freeform string.

Now we need to insert the raw flux measurements into the database. To do so, we will use the BrokerLc class that we created earlier. As an initial example, we will insert the first 3 observations from the source FAKE010.

There are two things to highlight above: (1) we do not specify an id for the data as this is automatically generated, and (2) the data insertion happens via a tuple. In this case, we are taking advantage of the fact that a Python tuple can be concatenated:

(objId,) + tuple(lc10.DFlc.ix[0]))

While the above example demonstrates the insertion of a single row to the database, it is far more efficient to bulk load the data. To do so we will delete, i.e. DROP, the rawPhot table and use some pandas manipulation to load the contents of an entire file at once via executemany.

Load all of the raw photometric observations into the rawPhot table in the database.

Problem 3b Load all of the raw photometric observations into the rawPhot table in the database.

Hint - you can use glob to select all of the files being loaded.

Hint 2 - you have already loaded the data from FAKE001 into the table.

Problem 3c

To ensure the data have been loaded properly, select the $g'$ light curve for source FAKE010 from the rawPhot table and plot the results. Does it match the plot from 1c?

While we cannot load the light curves to the database in parallel, using the concept of MapReduce we can use the iPython "cluster" that we created to map the I/O process (typically lots of overhead) to multiple machines, and then we can reduce their output to a single object to be loaded into the database.

Of course, before you parallelize, you profile.

Problem 3d

Earlier it was claimed that bulk loading the data is faster than loading it line by line. Prove this assertion, use %%timeit to "profile" the two different options (bulk load with executemany and loading one photometric measurement at a time via for loop).

Compare both serial solutions, to load the data in parallel.

Hint - to avoid corruption of your current working database, MiniBroker.db, create temporary databases for the pupose of running this test. Note - the creation and connection to the database should happen outside your %%timeit magic function. This means your solution will have (at least) 4 separate cells.

You shold have confirmed that loading a full light curve is faster than loading individual observations (not by a ton, but faster is faster and microseconds count when you are dealing with 37 billion sources).

Now we will parallelize this operation in order to perform the I/O in parallel, followed by a single bulk load of the data. To do so, we need the @require decorator, which will ensure the necessary packages and definitions are available on each computer within our cluster.

Now, we use the previously defined lview variable to achieve DirectView access to the cluster to map the load_data function on to all the filenames, and get the results.

And now that you have results, drop the rawPhot table if it already exists, and load the observations into the database as before.

Problem 3e Confirm that reading the data in parallel is faster than the seriel operations in 3d using the %%timeit function as before.

That isn't quite a factor of 4x speed up, but it represents a significant improvement, and recall that writing to the database must happen as a serial process because we are using SQLite.

Problem 3f

Once again -- ensure the data have been loaded properly by plotting the $g'$ light curve for source FAKE010 from the rawPhot table and plot the results. Does it match your earlier plots?

Problem 3) Part 2 -- Calculate and store the features in the same database

Now that we have loaded the raw observations, we need to create a new table to store summary statistics for each object. This table will include the Variable class attributes.

The above procedure should look familiar to above, with one exception: the addition of the foreign key in the lcFeats table. The inclusion of the foreign key ensures a connected relationship between rawPhot and lcFeats. In brief, a row cannot be inserted into lcFeats unless a corresponding row, i.e. objId, exists in rawPhot. Additionally, rows in rawPhot cannot be deleted if there are dependent rows in lcFeats.

Problem 3g

In parallel - calculate all the BrokerLc features for every source in rawPhot. Then, bulkload these features into the lcFeats table. Use the same map reduce procedure as above. When calculating the features create a tuple called feats with entries corresponding to the lcFeats table, in the exact order of the columns in that table.

Hint - it might be easier to write a seriel solution first and then introduce the DirectView implemenation. Feature calculation requires significant resources, so accomplishing this in parallel will save lots of time.

Hint 2 - technically, LSST light curves will be "generated" from the rawPhot table. However, skip that step for now as the BrokerLc class reads in data from a file, so loop over those files to populate the lcFeats table.

Problem 3h

Confirm that the data loaded correctly by counting the number of sources with gAmp > 2.

How many sources have gMinusR = -999?

Hint - you should find 9 and 2, respectively.

Finally, we close by commiting the changes we made to the database.

Note that strictly speaking this is not needed, however, were we to update any values in the database then we would need to commit those changes.

Problem 4) Build a Classification Model

One of the primary goals for LSST Brokers is to separate the Wheat from the Chaff. In other words, given that ~10 million alerts will be issued by LSST on a nightly basis, what is the single (or 10, or 100) most interesting alerts.

Here we will build on the skills developed during the previous homework to construct a machine-learning model to classify new light curves.

Fortunately - the data that has already been loaded to MiniBroker.db is a suitable training set for the classifier (we simply haven't provided you with labels just yet). Execute the cell below to add a new table to the database which includes the appropriate labels.

For now - don't worry about what the labels mean (though if you inspect the light curves you may be able to figure this out...)

Problem 4a

Query the database to select features and labels for the light curves in your training set. Store the results of these queries in numpy arrays, X and y, respectively, which are suitable for the various scikit-learn machine learning algorithms.

Hint - recall that databases do not store ordered results.

Hint 2 - recall that scikit-learn expects y to be a 1d array. You will likely need to convert a 2d array to 1d.

Problem 4b

Train a SVM model (SVC in scikit-learn) using a radial basis function (RBF) kernel with penalty parameter, $C = 1$, and kernel coefficient, $\gamma = 0.1$.

Evaluate the accuracy of the model via $k = 5$ fold cross validation.

Hint - you may find the cross_val_score module helpful.

The SVM model does a decent job of classifying the data. However - we are going to have 10 million alerts every night. Therefore, we need something that runs quickly. For most ML models the training step is slow, while predictions are (relatively) fast.

Problem 4c

Pick any other classification model from scikit-learn, and "profile" the time it takes to train that model vs. the time it takes to train an SVM model (by using the %%timeit magic function).

Is the model that you have selected faster than SVM?

Hint - you should import the model outside your timing loop as we only care about the training step in this case.

Problem 4d

Does the model you selected perform better than the SVM model? Perform a $k = 5$ fold cross validation to determine which model provides superior accuracy.

Problem 4e

Which model are you going to use in your Broker? Justify your answer.

Write solution to 4e here

Problem 5) Class Predictions for New Sources

Now that we have developed a basic infrastructure for dealing with streaming data, we may reap the rewards of our efforts. We will use our ANTARES-like software to classify newly observed sources.

Problem 5a

Load the light curves for the new observations (found in full_testset_for_LSST_DSFP) into the a table in the database.

Hint - ultimately it doesn't matter much one way or another, but you may choose to keep new observations in a table separate from the training data. I put these observations into a new testPhot database. Up to you.

Problem 5b

Calculate features for the new observations and insert those features into a database table.

Hint - again, you may want to create a new table for this, up to you. I'm using the testFeats table.

Problem 5c

Train the model that you adopted in 4e on the training set, and produce predictions for the newly observed sources.

What is the class distribution for the newly detected sources?

Hint - the training set was constructed to have a nearly uniform class distribution, that may not be the case for the actual observed distribution of sources.

It turns out that the classification performance in this case isn't particularly good.

For the test set, the actual class labels were sneakily hidden in the file headers, and if you execute the cell below you will discover how many sources are in each class.

As we discussed in the lecture, one challenge for this is problem is that individual observations are largely spaced out, and have very large uncertainties. The features that we have currently adopted (which are really just moments of the flux distribution) aren't robust to outliers, and are subject to the precise cadence of the observations.

One way to "correct" for the observational gaps and heteroskedastic uncertainties is to interpolate a model between the observations. A method that has become really popular recently is the use of Gaussian process regression (there will be more along these lines in the next homework assignment, but here we will provide a preview).

Lets look at one fit to see what's happening:

Not only does GP regression provide an interpolation between observations, but it returns an uncertainty about that interpolation as well. Thus, it's possible to measure features on interpolated light curves that are observed with precisely the same cadence for all sources.

But here's the rub - can we practically run this on 10^4--5 alerts every 37 seconds for LSST?

This is one reason parallelization is so important - as long as we have enough computers, we can easily keep up with the LSST data rate even with complex features (e.g., periodicity search, GP regression, etc).

Problem 6) Anomaly Detection

As we learned earlier - one of the primary goals of LSST brokers is to reduce the stream of 10 million alerts on any given night to the single (or 10, or 100) most interesting objects. One possible definition of "interesting" is rarity - in which case it would be useful to add some form of anomaly detection to the pipeline. scikit-learn has several different algorithms that can be used for anomaly detection. Here we will employ isolation forest which has many parallels to random forests, which we have previously learned about.

In brief, isolation forest builds an ensemble of decision trees where the splitting parameter in each node of the tree is selected randomly. In each tree the number of branches necessary to isolate each source is measured - outlier sources will, on average, require fewer splittings to be isolated than sources in high-density regions of the feature space. Averaging the number of branchings over many trees results in a relative ranking of the anomalousness (yes, I just made up a word) of each source.

Problem 6a

Using IsolationForest in sklearn.ensemble - determine the 10 most isolated sources in the data set.

Hint - for IsolationForest you will want to use the decision_function() method rather than predict_proba(), which is what we have previously used with sklearn.ensemble models to get relative rankings from the model.

Problem 6b

Plot the light curves of the 2 most anomalous sources.

Can you identify why these sources have been selected as outliers? Are these sources in fact outliers?

Write solution to 6b here

Problem 7) Real Life

Problem 7a

We have made many simplifying approximations throughout this notebook. What do you think are the most difficult challenges when extending the "toy problem" examples here to a real broker for LSST alerts?

Write solution to 7a here