IPP to PSPS interface: ippToPsps

ippToPsps is the interface between IPP and PSPS. At the highest level, its job is to create FITS files, generated from a multitude of IPP data-sources, and then publish them to a datastore in the form of batches. On the PSPS side, the DXLayer polls the datastore, collects batches when they become available, then converts the contents to csv files before sending them on to SQL Server loader software, which merges them into the relevant PSPS database.

The binary tables in the FITS files generated by ippToPsps match the PSPS database schema perfectly, the consequence being that any alterations to the PSPS database schema will only affect ippToPsps code, and not the DXLayer. A certain amount of data validation is performed by ippToPsps before publication, with more validation occurring at the loading and merge stages on the PSPS side.

Simultaneously to loading data, ippToPsps polls PSPS to inquire after the status of the batches it has loaded. Batches that failed can then be reloaded. The status of every batch is maintained in a MySQL database local to ippToPsps. The various temporary copies of the batches are deleted once the ODM reports that they have been successfully loaded or merged.

Some notes prepared for the PSPS operational readiness review can be found here.

Batch types

The outputs of ippToPsps are referred to as 'batches', and are detailed below.

Batch name	PSPS name	Description	IPP Source
Initialization	IN	metadata relating to other batches, eg filter ID, survey ID etc	generated from XML config
Detections	P2	single exposure detections	generated from one smf file per exposure plus associated DVO database
Stack	ST	stack image detections	one FITS file generated per IPP cmf, which contains data for one filter on one skycell

Architecture and design

Outline

The job of ippToPsps is to take a collection of IPP tables and convert them to tables suitable for ingestion to PSPS. Some mappings between IPP and PSPS are direct (eg IPP exposure ID = PSPS frame ID), while many require a format conversion, or can only be derived from multiple IPP fields; some require mining info from other IPP sources, such as DVO databases or the IPP MySQL database, gpc1.

Languages and tools used

The tools chosen for ippToPsps are those considered to be the most effective to tackle the task at hand. As a result, ippToPsps is not consistant with the rest of the IPP code-base (which is predominately Perl and C), but since its role is essentially outside the IPP, this, to me, is not an issue.

MySQL

Because we are dealing with table data that requires some intensive manipulation, it follows that a relational database be used. Relational databases are highly optimized to provide extremely fast query times, especially when indexing in incorporated. Thus we gain speed over the more obvious route of read-a-table-from-FITS-into-an-array-then-loop-through-each-value etc. We also have to write far fewer lines of code.

An added bonus is that, by using keys that enforce uniqueness in a given column (or columns), we protect ourselves against the risk of duplicates making it into PSPS.

ippToPsps uses two MySQL databases, a scratch database, used to import tables and manipulate then before discarding them, and the ippToPsps database, which keeps track of which batches have been processed, published to PSPS etc.

Jython and STILTS

ippToPsps is written in ​Jython, this is in part to take full advantage of the ​STILTS package, which enables fast and efficient processing of astronomical catalog data tables. Since it supports FITS, VOTable, and SQL, it is a perfect fit for this project. It is also software maintained elsewhere, reducing the burden on us.

Jython is simply a Java implementation of Python, a modern, high-level, object-oriented language enabling ippToPsps to be written in minimal lines of code, helping it be both more readable and maintainable.

High-level design

Each batch type produced by ippToPsps (detection, stack, init etc) has its own class, all of which inherit from the Batch base-class, which handles features common to all batches, such as creating and opening output FITS files, connecting to the GPC1 database, connecting with the DVO database etc. This keeps duplicated code at a minimum. The Batch class is an abstract class, i.e. it should not, and cannot, be instantiated.

ippToPsps works like this:

• reads all relevant FITS tables from a given smf or cmf into temporary tables in a 'scratch' MySQL database
• creates empty MySQL tables for PSPS output (also in the 'scratch' database). These tables match the shape of the final PSPS database tables exactly.
• copies all relevant columns from the temporary IPP tables into the PSPS tables, discarding duplicates where necessary
• accesses the DVO database and creates temporary MySQL tables containing all the detections for this smf or cmf
• updates the PSPS tables with the IDs from the DVO MySQL tables
• replaces any NULL values with -999, as required by the PSPS loader
• exports the PSPS tables to a FITS file
• publishes FITS file, complete with a batch 'manifest' file, to the datastore

The reading and export of FITS tables is done using STILTS. For import, we can specify which columns we wish to import from the IPP smf and cmf files (we don't need everything).

Configuration

Due to the potential for changes in both input and output for ippToPsps, rather than hard-coding table descriptions, the code is heavily configurable. As such, the tables descriptions are stored as ​VOTable files (for ​example), which are a standard of the ​IVOA. Because VOTables are an XML format, they are both human and machine readable, expandable and self-describing. These VOTables are generated directly from the PSPS schema using a ​script, so that any changes to the schema can be easily passed-along to ippToPsps.

Another configuration file is used to describe a particular 'loading campaign', complete with the names of databases, location of output files, datastore settings etc. The path to one of these configuration files is usually the sole argument when running any of the programs described below. An example can be seen ​here.

Editing a config is easy, as it is XML format and most fields are self-explanatory. Others that are less obvious will be explained below in the Using the software section.

The ippToPsps database

The ippToPsps database is a central component. It is currently hosted on the same machine as gpc1, but all databases used by ippToPsps are detailed in the settings file, ​here, while it's schema can be found in svn, ​here.

dvoToMySQL

For the special case where we have a 'locked' DVO database, i.e. one that is complete and will not grow any further, we can improve ippToPsps processing time by first generating MySQL tables that include all detections from DVO rather than just the detections-per-exposure (or stack) as above. For this we can use the dvoToMySQL tool, found ​here

It may take a matter of days to convert a relatively small DVO database to MySQL, however, querying the MySQL database is hugely faster than accessing DVO directly, especially for regions of sky with a high density of detections such as the medium deep fields. (This was seen when loading MD4 prior to the Boston meeting in May 2011. DVO access per exposure was 40 minutes, whereas, once imported to MySQL, query time was roughly 30 seconds.)

Image ID confusion

We access DVO via a combination of 'source ID' and 'image ID'. Both numbers come from the smf file. However, IMAGE_ID in the smf does not correspond to IMAGE_ID in DVO, instead, it maps to EXTERN_ID in the Images.dat file at the top-level of a given DVO database. The IMAGE_ID column of the same table maps instead to IMAGE_ID in the various 'cpm' (measurements) files contained within the subdirectories of the same DVO database.

Using the software

ippToPsps contains numerous separate programs that work together to process and load IPP data (in the form of batches) to the datastore, monitor the progress of those batches and then perform the necessary cleanup of the various temporary files once those batches are safely merged into the PSPS database.

Loading

Loading data from IPP to PSPS means running an instance of the load.py program. It is possible, and usually necessary, to run multiple versions of load.py. The code has been designed so that the user can start as many clients as he or she likes, and they will run concurrently without causing any problems. Note, however, that clients on machines where the DVO database is not on a local disk will run more slowly due to the increased burden of accessing DVO over the network.

Any number of clients can run concurrently on any number of hosts, so long as they have a local 'scratch' database set up properly, as detailed below.

Setting up a loading machine

Although this can be changed in the config file, ippToPsps uses a 'scratch' MySQL database on localhost while processing. So, to to any loading, a local MySQL server should be up-and-running, with a database created as the ipp user for use by ippToPsps. The default name for this is ipptopsps_scratch.

Because ippToPsps uses this local database as a scratch database, there are large number of inserts and deletions, meaning large log files. For this reason, it is important to configure MySQL to store its data on a local partition, and not /var as is the default. For example, on ipp005:

/var/lib/mysql

was moved to:

/export/ipp005.0/mysql

In addition, it is prudent to decrease the default MySQL log-retention period from 90 days to 1 day. This is done is in

/etc/mysql/my.cnf 

on the line

expire_logs_days                                = 90

You will need (if it is not done already) to create an ipp user. Connect to the local MySQL server as root then to the following.

GRANT ALL PRIVILEGES ON *.* TO 'ipp'@'localhost' IDENTIFIED BY 'ipp' WITH GRANT OPTION;

Now disconnect then reconnect as ipp and create the scratch database.

create database ipptopsps_scratch;

Finally, you need to create the initialization tables.

./run.sh load.py configs/init.xml init

At the time of writing, scratch databases are set up on the following IPP hosts:

ipp005
ipp006
ipp007
ipp033
ipp034
ipp045

ipp005 and ipp045 in fact have three scratch databases each, so three clients can potentially be run on these machines. This is particularly useful on ipp005 as this is where the DVO databases are currently stored, so clients running on the same machine have a much higher throughput. (ipp045 used to host a backup 3pi DVO database.)

Running

Loading data means running the load.py program while passing it the relevant configuration file (see above). So, running the software is simply a case of editing a config and passing it as an argument to the program, like this:

cd trunk/ippToPsps/jython
./run.sh load.py someConfig.xml

The ./run.sh prefix above is necessary to invoke the correct Java virtual machine, while including the relevant jar files in the CLASSPATH (all included in the trunk/ippToPsps/Jars subdir)

Creating and publishing an 'IN' batch

The same program is used to create IN batches, but requires one extra argument, like this:

./run.sh load.py someConfig.xml init

This will produce an IN batch, publish it, then stop (the IN-batch MySQL tables will also be updateded in the local scratch database).

Queuing

When run, load.py first queues up all the exposures or stacks that it needs to process, then works through them sequentially, doing detections first then stacks. The master list comes from the gpc1 addRun table, which lists all items currently merged into the DVO database we are using. From this list, ippToPsps subtracts items that have already been successfully loaded to the datastore already, and queues up the remainder.

Because it is possible, and usual, to run multiple versions of load.py in order to speed up loading time, the methods to queue and begin processing a new batch form a critical section. This simply means that the 'batch' table in the ippToPsps database is locked by a client looking for a new item to process, then released afterwards.

An important factor in the queuing of items is the current epoch, which is set in the config file.

Epochs

Each config file requires an entry in the options/epoch field. An epoch in the context of ippToPsps is the date that we count as the beginning. If we loaded IPP data to PSPS once and only once, this would not be necessary: we would queue up all available exposures and stacks and simply load them. However, the early stages of the project have required multiple re-loads of data while the IPP perfects the science, necessitating that the same exposures and stacks are loaded more than once. Because load.py queues available items by taking into account those that have already been published to PSPS, we have to give it an epoch date from which to accurately determine those exposure that have been loaded or not. Generally speaking, the epoch is reset every time we delete the main PSPS database.

Test mode

One field in the config, under the options section, is testMode. This is a boolean, i.e. can be either 1 or 0. When switched on, each batch type will be run in it's own test mode. For detections, this means:

• it will only process the first exposure that is queued
• it will only process OTA 33 from this exposure
• it will ignore certain missing values, which would otherwise fail the batch

Running dvoToMySQL

Like all other ippToPsps components, dvoToMySQL takes one argument which is a path to a config file. For example:

./run.sh ./dvoToMySQL.py configs/lap.xml

This will begin to import the DVO database pointed to in the config into the scratch MySQL database also detailed in the config. The program can be started and stopped, as it will 'remember' which DVO files it has read and imported.

There is a second 'hidden' argument to dvoToMySQL, namely reset. This is dangerous, as this option will reset all the tables necessary to import a DVO database. If you have spent days importing a DVO database, then run dvoToMySQL using this option, everything wilol be lost. The user is prompted to ask whether they really want to proceed.

Deletion policy

When loading, three copies of the data exist: the original files on disk, the batches on the datastore and a third copy on the PSPS loading machine. The deletion policy at present is:

• when a batch has been loaded to the ODM and has a status of 'merge worthy' or 'failed', then the copy on the datastore and the DXLayer's copy can be removed
• when a batch has been successfully merged into the PSPS database, then final copy on local disk may be deleted

The logic for this is that errors may occur during the merge phase and it is useful to have local copies of offending batches for debugging purposes. This standard behavior can be changed by setting the appropriate values in the deletion section of the config file.

To actually perform the clean, run the cleanup.py program like this:

./run.sh cleanup.py  configs/lap20110809.xml

Metrics

ippToPsps has a metrics program which reports processing and loading progress per batch type. It can be run, for example, like this:

./run.sh metrics.py configs/lap20110809.xml

And will produce a response similar to this:

                ippToPsps loading summary

Time now                                2011-09-29 11:54:33
Loading epoch                           2011-08-16
DVO label                               LAP.ThreePi.20110809

+----+------------------+---------------+-------------------+------------------+----------------+
|Type| batches per hour | last 24 hours | per day this week | total detections | last published |
+----+------------------+---------------+-------------------+------------------+----------------+
| P2 |              0.0 |           283 |             329.4 |       1855730051 |  4.6 hours ago |
| ST |              0.0 |         12936 |            4539.0 |        185807608 |  3.7 hours ago |
+----+------------------+---------------+-------------------+------------------+----------------+

+----+-------+-------------------+-------------------+-------------------+-------------------+-------------------+-------------------+-------------------+-------------------+
|Type|  DVO  |          processed|loaded_to_datastore|      loaded_to_ODM|       merge_worthy|             merged|  deleted_datastore|    deleted_dxlayer|      deleted_local|
|    |       | Pend  Succ  Fail  | Pend  Succ  Fail  | Pend  Succ  Fail  | Pend  Succ  Fail  | Pend  Succ  Fail  | Pend  Succ  Fail  | Pend  Succ  Fail  | Pend  Succ  Fail  |
+----+-------+-------------------+-------------------+-------------------+-------------------+-------------------+-------------------+-------------------+-------------------+
| P2 |  8140 |       8101    39  |    2  8099        | 2041  6049     9  |       6049        |  313  5736        |       5736        | 5736              |                   |
| ST | 32309 |      31684   625  |    5 31679        |24617  7062        |       7062        | 7062              |                   |                   |                   |
+----+-------+-------------------+-------------------+-------------------+-------------------+-------------------+-------------------+-------------------+-------------------+

It will also produce plots in the jython/plots directory, such as this:

Recovery system design

Currently, the IPP to PSPS interface is a 'one-way' system. Batches are created by ippToPsps and posted on an IPP instance of the datastore. These batches are collected by the DXLayer on the PSPS side and sent on to the ODM. The IPP urgently requires some feedback from PSPS to determine which batches have succeeded and which have failed (and why they failed). With this information, data can be either deleted or regenerated accordingly. This is important simply because, with such large data volumes, we cannot afford the high levels of redundancy currently in place. At present, for a given batch, the following copies exist within the pipeline:

• a copy exists on the IPP cluster after generation by ippToPsps program
• a copy exists on the IPP datastore after publication by ippToPsps
• the DXLayer retains a copy after it has sent the csv version to the ODM
• the DXLayer also keeps a copy of these (larger) csv files

We therefore need to quickly implement the basic framework of a feedback loop so that the IPP can quickly learn if a given batch has been successfully merged into the PSPS database or not. This will enable it to safely delete the data files and remove the copy from the datastore. This will also form the basis for a more comprehensive recovery system, to be developed at a future date.

Previous design

                                              .............................
                                             .   ___________               .
                                             .  |           |              .
         ---------------------------------------| datastore |              .
        |                                    .  |___________|              .
 .......|...........................         .       /|\                   .
.  ____\|/_____       ___________   .        .    ____|____        _____   .
. |           |      |           |  .        .   |         |      |     |  .
. | ippToPsps |----->| datastore |-------------->| DXLayer |<---->| ODM |  .
. |___________|      |___________|  .        .   |_________|      |_____|  .
.                                   .        .                             .
 ...................................         ..............................   
                
               IPP                                       PSPS                         
                                         

Previously, Conrad and I had discussed a design whereby a second datastore instance would be utilized, this time on the PSPS cluster. The DXLayer would act as the 'middle-man', polling the ODM for updates on loading progress, then posting the results on the PSPS datastore for the IPP to consume. Polling this, ippToPsps could acquire a list of batches it knows are safe to be discarded. Simultaneously, the DXLayer could delete its copies of the same redundant data.

The update placed on the PSPS datastore could take the form of an XML file. At first this would simply detail those files it is safe to delete, but could evolve into a more complex recovery report, i.e. which batches failed, and what is required to be done by the IPP.

New design

                                                                                                      
         ------------------------------------------------------------        
        |                                                            |     
........|...........................          .......................|.....                 
.  ____\|/_____       ___________   .        .    _________        __|__   .
. |           |      |           |  .        .   |         |      |     |  .
. | ippToPsps |----->| datastore |-------------->| DXLayer |----->| ODM |  .
. |___________|      |___________|  .        .   |_________|      |_____|  .
.                                   .        .                             .
 ...................................         ..............................   
                
               IPP                                       PSPS                         
                                         

Instead of creating a new datastore instance within PSPS and using the DXLayer as communication layer between the ODM and the IPP, we propose that the DXLayer forms no part of the feedback system. It should be simplified such that it only facilitates loading, i.e. polling the IPP datastore for new data, converting it to csv files then sending these on to the ODM. Instead, to complete the circle, the ippToPsps code will poll the ODM directly, bypassing the DXLayer altogether. This also forms the basis of a full recovery system as, at a later date, ippToPsps can be coded to respond intelligently to the myriad of errors that may occur within the ODM. The DXLayer need know nothing of the how or why a certain batch is being submitted by the IPP, it should just grab it, convert it and pass it along to the ODM.

This design would therefore mean simplifying a major PSPS component, the DXLayer, but rather than waste the code already written, it would be taken and used within ippToPsps (for example, the ODM polling scripts). We would simply be shifting responsibility over from PSPS to IPP. Over parts could be dropped completely. For example, since ippToPsps will (soon) keep a record of all the jobs and corresponding exposure IDs in the IPP database, it is unnecessary for this information to be duplicated by the DXLayer, which currently has its own local database for this information.

The question remains of what should be done with the copies of the data currently retained by the DXLayer? The options are that it can either be deleted automatically after a defined amount of time, or the IPP can send a list of batches it is safe to delete through the datastore, or perhaps the DXLayer should not retain files at all. Since it can quickly and easily acquire data from the IPP datastore anyway, it is probably unnecessary for it to hold any copies.

Advantages over previous design

• no need for second datastore (not a big overhead, but it would require additional systems administration in an already complicated system)
• no need to define new XML standard that incorporates the whole array of recovery options
• no need for the DXLayer to poll the ODM
• no need for the DXLayer to have a database to log the batches (already done on the IPP side)
• no need for the DXLayer to keep data at all?

Links

​Datastore test area for PSPS on Maui
​Datastore test area for PSPS at JHU