|Cooking with Sloan|
|Query Form Guide|
|SQL in SkyServer|
|Sample SQL Queries|
|Contact Help Desk|
SQL is the Structured Query Language, a standard means of asking for data from databases, and is used to query the Catalog Archive Server (CAS). This page provides a brief overview of SQL. Query examples are also available, with comments, as well as a page of links to more detailed off-site documentation.
When performing queries, you must first decide which database you will be using. There are two main databases in the CAS, Target and Best. In the DR4, these databases are actually named TARGDR4 and BESTDR4. The Target database contains all measurements as they were made when objects were targeted for spectroscopy. Best contains the best data and most recent processings for the entire released sky area. The area coverage is almost, but not exactly, the same. By default, queries are made on the Best database. To use a different database, you can use the .. syntax to specify a table in the other database, for instance:
For more details on the differences between Target and Best, please see the data model page.
Each database contains a large number of tables, some of which contain photometric measurements (such as PhotoObj), spectroscopic measurements (such as SpecObj), or information about the observing conditions (Field) or survey geometry(TileBoundary). See the data model page for more details.
In addition to the tables, we have defined Views, which are subsets or combinations of the data stored in the tables. Views are queried the same way Tables are; they exist just to make your life easier. For instance, the view Galaxy can be used to get photometric data on objects we classify as galaxies, without having to specify the classification in your query.
Both the Skyserver and CasJobs interfaces have a Schema Browser. It shows you all of the available databases, the tables in each database, and the quantities stored in each column of the tables.
Finally, we have created a variety of functions and stored procedures which let you easily perform some common operations. Usually, their names are prefixed by f or sp, like in fPhotoStatus or spGetFiberList. The full list of functions and store procedures is found in the Schema Browser. Note that some functions are scalar-valued, meaning that they return a single value, while others (such as the commonly used dbo.fGetNearbyObjEq, are table-valued; they actually return a table of data, and not a single number. This is important when interpreting the returned data and performing joins.Please note the caution about using function calls as noted in the Optimizing Queries section when attempting queries over that return a large number of objects.
Now that we have an overview of the database structure, how do we actually get data out? You will have to write a query using SQL. The most basic query consists of three parts:
The WHERE clause is not necessary if you want to retrieve parameters of all objects in a specified table, but this typically will be an overwhelming amount of data!
Note that the query language is insensitive to splitting the query over many lines. It is also not case sensitive. To make queries more readable, it is common practice to write the distinct query clauses on separate lines. The Sample Queries button on the CasJobs Query page provides a variety of samples, ordered in complexity. For instance, to obtain the list of unique Fields that have been loaded into the database, we use:
SELECT FieldID FROM Field
If we want to retrieve multiple parameters from the database, we separate them with commas:
SELECT ra,dec FROM Galaxy
Of course, the parameters you request must be included in the table(s) you are querying! Now, let's say we want magnitudes of all bright galaxies. We will need to specify a magnitude range to do this:
SELECT u,g,r,i,z FROM Galaxy WHERE r<12 and r>0
Here, we have used the WHERE clause to provide a magnitude range. The and operator is used to require that multiple limits be met. This leads us to...
Simple Logical and Mathematical Operators
Not only can we place limits on individual parameters, we can place multiple limits using logical operators, as well as place limits on the results of mathematical operations on multiple parameters. We may also retrieve results that are logical joins of multiple queries. Here we list the logical, comparison, and mathematical operators.
The LOGICAL operators are AND,OR,NOT; they work as follows:
When comparing values, you will use the COMPARISON operators:
In addition to the comparison operators, the special BETWEEN construct is available.
Finally, the MATHEMATICAL operators (both numeric and bitwise) are:
In addition, the usual mathematical and trigonometric functions are available in SQL, such as COS, SIN, TAN, ACOS, etc..
Querying Bit Flags
Several SDSS tables contain bit-encoded flags to indicate various types of information about the object or quantity in question (e.g., PhotoTag and PhotoObjAll tables each have the flags field, SpecObj has zWarning flags etc.). This section describes how you can test for flag values in your query. For sample queries that demonstrate the use of flags, see the Errors using flags, Elliptical galaxies with model fits, Diameter limited sample, LRG sample, and Clean photometry with flags sample queries for examples on how to use flags.
Checking a single flag
To return rows for which the flag is set, the basic syntax for the constraint is:
and to return rows for which the flag is not set:
where bitmask is the binary value in which the bit corresponding to the flag is 1 and all other bits are 0. You can use the flag functions provided by the SkyServer (listed in Schema Browser) to get the bitmask for a given flag, but if you are scanning a large fraction of a large table like PhotoObj, you are better off not making a function call for each row, and in that case you should first get the actual binary value of the bitmask first and substitute that instead. This is described in the Using dbo functions in your query subsection of the Optimizing Queries section below.
For example, to select objects for which the BLENDED flag is set in PhotoTag, you would use a query like:
and to select only objects for which the flag is NOT set, use
Checking multiple flagsTo test if multiple flags are set, you can combine the values by adding them and then testing the result.
To select objects for which all of several flags are set, generate the combined bitmask by adding the individual flag bitmasks, then compare the result of ANDing the combined bitmask with the flag column with the combined bitmask itself, e.g.,
To select objects for which at least one of several flags is set, you just need to check that ANDing the combined bitmask with the flag column returns a non-zero result, e.g.,
To select objects for which none of several flags is set, the result of ANDing the flag column with the combined bitmask must be 0, e.g.,
The SDSS photo pipeline sets a number of flags that indicate the quality of the photometry for a given object in the catalog. If you desire objects with only clean photometry for science, you should be aware that you need to filter out unwanted objects yourself in your query. This is not done automatically for you (e.g. with a view of the PhotoObjAll table). The main reason is that the flag constraints that are required for this filtering often impose a significant performance penalty on your query, and may even invoke the bookmark lookup bug.
Please see the Clean Photometry sample query for help on how to use the photometry
flags to select only objects with clean photometry.
Excluding Invalid Data Values
As mentioned in the EDR Paper, the database designates quantities that are not calculated for a particular object in a table with special values, as follows:
Changing Precision of Query Output
Use the STR(column,n,d) SQL construct (where n is the total number of digits and d is the number of decimal places) to set the precision of the column that your query requests. The SkyServer returns values with a default precision that is set for each data type, and this may not be enough for columns like ra, dec etc. See the Selected neighbors in run or the Uniform Quasar Sample sample queries for examples of how to use STR.
Joins: Querying With Multiple Tables
You may wish to obtain quantities from multiple tables, or place constraints on quantities in one table while obtaining measurements from another. For instance, you may want magnitudes (from PhotoObj) from all objects spectroscopically identified (SpecObj) as galaxies. To perform these types of queries, you must use a join. You can join any two (or more) tables in the databases as long as they have some quantity in common (typically an object or field ID). To actually perform the join, you must have a constraint in the WHERE clause of your query forcing the common quantity to be equal in the two tables.Here is an example, getting the g magnitudes for stars in fields where the PSF fitting worked well:
Notice how we define abbreviations for the table names in the FROM clause; this is not necessary but makes for a lot less typing. Also, you do not have to ask for quantities to be returned from all the tables. You must specify all the tables on which you place constraints (including the join) in the FROM clause, but you can use any subset of these tables in the SELECT. If you use more than two tables, they do not all need to be joined on the same quantity. For instance, this three way join is perfectly acceptable:
The type of joins shown above are called inner joins. In the above examples, we only return those objects which are matched between the multiple tables. If we want to include all rows of one of the tables, regardless of whether or not they are matched to another table, we must perform an outer join. One example is to get photometric data for all objects, while getting the spectroscopic data for those objects that have spectroscopy.
In the example below, we perform a left outer join, which means that we will get all entries (regardless of matching) from the table on the left side of the join. In the example below, the join is on P.objID = s.BestObjID; therefore, we will get all photometric (P) objects, with data from the spectroscopy if it exists. If there is no spectroscopic data for an object, we'll still get the photometric measurements but have nulls for the corresponding cpectroscopy.
You can join across more than one table, as long as every pair you are joining has a quantity in common; not all tables need be joined on the same quantity. For example:
Note how the Field and PhotoObj are joined on the fieldID, while the join between Field and Segment uses segmentID.
When using table valued functions, you must do the join explicitly (rather than using "="). To do this, we use the syntax
Manipulating Query Output
SQL provides a number of ways to reorder, group, or otherwise arrange the output of your queries. Some of these options are:
It is easy to construct very complex queries which can take a long time to execute. When writing queries, one can often rewrite them to run faster. This is called optimization.
The first, and most trivial, optimization trick is to use the minimal Table or View for your query. For instance, if all you care about are galaxies, use the Galaxy view in your FROM clause, instead of PhotoObj. We have also created a 'miniature' version of PhotoObjAll, called PhotoTag. This miniature contains all the objects in PhotoObjAll, but only a subset of the measured quantities. Using the PhotoTag table to speed up the query only makes sense if you do NOT want parameters that are only available in the full PhotoObjAll.
It is extremely useful to think about how a database handles queries, rather than trying to write a plain, sequential list of constraints. NOT every query that is syntactically correct will necessarily be efficient; the built-in query optimizer is not perfect! Thus, writing queries such that they use the tricks below can produce significant speed improvements.
Here is a staggering example of the importance of optimization:
A user's first instinct would be to get the desired objects from the PhotoObj table within the TARGDR4 database (which contains the information, including targeting decisions, for objects when they were targeted (chosen) for spectroscopy). So, this query might look like:
That's really simple - all you are doing is checking if the primary target flags (primtarget) are set for the two types of QSO targets. This query can take hours, because a sequential scan of every object in the photometric database is required!
One quick change which makes a difference is to simplify the WHERE clause, to get rid of the or, by masking everything but bits 2,4, and checking if the result is non zero. This changes the WHERE clause to:
This helps a little, but not much - we are still scanning the entire PhotoObj table. We can make our lives a lot better by realizing that the database developers have anticipated that people will be interested in targetting information, and created a smaller table TargetInfo, that contains only the Targetted objects, which is a small subset of the entire photometric database! Using this table, we can rewrite our query as (we do need a join with the Target table also):
Note how most of the WHERE clause is performed using the Targetinfo table; the SQL optimizer immediately recognizes that this table is much smaller than PhotoObj, and does this part of the search first. The query now runs in about a minute or two. That is two orders of magnitude improvement over the initial method!.
Finally, we can recognize that all the quantities of interest are also in the PhotoTag table, which contains all the objects in PhotoObjAll, but not all measured quantities. The query will be:
This runs in 18 sec, and returns the same 32931 rows. Another factor of two in speed! Note how PhotoTag does not contain the simplified i magnitude, and we must use ModelMag_i instead.
Another of the simplest ways to make queries faster is to first perform a query using only indexed quantities, and then select those parameters from the returned subset of objects. An indexed quantity is one where a look-up table has effectively been calculated, so that the database software does not have to do a time-consuming sequential search through all the objects in the table. For instance, sky coordinates cx,cy,cz are indexed using a Hierarchical Triangular Mesh (HTM). So, you can make a query faster by rewriting it such that it is nested; the inner query grabs the entire row for objects of interest based on the indexed quantities, while the outer query then gets the specific quantities desired.
Using Views is convenient, but ...This is a cautionary note about using views. Views are a great convenience that enable you to access virtual subsets of tables. The Archive Intro page lists the views defined on each table in the SDSS databases. However, there are a couple of points to remember with regard to using views, when it comes to performance:
Using views of PhotoObjAll vs PhotoTag tableWe've already seen above how using the PhotoTag table speeds up queries, but this point needs reiterating, especially in the context of queries that search a large fraction of the database. For queries that do not search large subsets of the DB (large subset = significant fraction of spatial coverage, that corresponds to millions of rows), it is convenient and recommended that you use the PhotoObjAll views (PhotoObj, PhotoPrimary, Star, Galaxy, etc.). However, for large queries that search over a good fraction of the database, there are at least 2 reasons why you should use the PhotoTag table instead in spite of the inconvenience of having to explicitly specify the mode (primary/secondary) and/or the object type (star, galaxy etc.) in the WHERE clause:
This query invokes the bookmark lookup bug and takes several hours to run as it is written above, if the TOP 100 is removed. If it is rewritten to use the PhotoTag table twice, instead of the Star view and the PhotoObj view, then it does not do a bookmark lookup and runs in a little over 2 hours on DR2. Note that the shorthand u,g,r,i,z magnitudes and associated errors have to be replaced with the full modelMag names since PhotoTag does not have the shorthand magnitudes.
Using dbo functions in your queryFinally, a caution about using function calls in queries. If your query is going to match a large number of objects (million or more), using a function call, especially one that operates on a constant or literal, in the WHERE clause is not a good idea, because the function will be called once per matching row in that table, resulting in a significant performance hit. Here is an example of this:
In this case, it would be better to first do the pre-query:
to get the bitmask value for that flag, and then rewrite the above query as:
This will avoid the wastefully repeated function call for each and every photobj in the table.
Performance and Indices
Performance is usually only an issue when the PhotoObjAll table (and associated views) is involved in a query, either directly or with a join. We have built in some features to enhance performance for queries on this table. The first and foremost, and the most effective performance enhancer, is the Hierarchical Triangular Mesh (HTM) spatial index that we have developed at JHU and incorporated into each of the SDSS databases. This is a multi-dimensional index that speeds up searches by spatial decomposition of the sky.
In addition to the HTM, there are several indices built in the database on
columns of the various tables, including primary key, foreign key and other
indices that group frequently used columns.
PhotoTag is a 10% subset of PhotoObjAll that has the 60 most "popular" fields.
Both PhotoObj and PhotoTag are indexed and those indices are each a 2% subset of PhotoObj.
The nice thing about the indices is that they get picked for you automatically and they run 50x faster than reading the whole PhotoObj table and 5x faster than reading the PhotoTag table.
The next version of the SQL Server database product will allow us to eliminate PhotoTag (it will be an automatically selected index). But for now, cognoscenti will have to use it if they can (if their question is covered by that 10% of the most popular fields).
In an ideal world you would not have to know about indices. Unfortunately we do not live in an ideal world (yet).
The strategy for selecting a few (less than 10,000) objects in a certain part of the sky using the dbo.fGetObjFromRect() function works very well. But, when the patch gets LARGE (more than 10,000 objects) then your ra-dec limit predicate is probably going to be more efficient because it will be a linear scan over the data.
The Stars/Galaxy/PhotoPrimary/... Views all benefit from the indices on the base tables. You should feel free to use them.
SQL Optimizer Bookmark Lookup Bug
The PhotoObjAll table is by far the largest table in the database, so queries on this table (and its allied views) are the only ones that will be too slow under certain circumstances. At current disk speeds (~ 400 MB/s peak), it should take about 15 minutes to do a sequential scan of the entire PhotoObjAll table in the BESTDR1 database (300+ GB), and about 30 minutes for BESTDR2..PhotoObjAll (700 GB), on an unloaded server. So even queries that scan the entire photoobj table should run in about half an hour if they are not requesting a very large number of rows (in which case it takes a long time to get the results back over the network).
Sometimes queries can run much slower than normal (5-10 times slower) if the server is loaded down, so you should always try a slow query at a few different times.
If after applying the advice given above and trying your best to optimize your query, you find that it still runs very slowly (no output returned in more than an hour or so, or your query times out), you may have run into the dreaded bookmark lookup bug of the SQL Server query optimizer. Basically, this means that the optimizer has chosen the incorrect plan for executing the query.
While there is no reliable way to predict what causes the bookmark bug to be invoked, usually it happens when there are several constraints on non-indexed quantities in a given table. For example, in the query
If you include only one of the two constraints separated by the "and", the optimizer will chose the correct plan, but if you include both, the optimizer goofs big-time and opts to do a random search of the PhotoObj table instead of a sequential scan. It decides that it will use the photoobj primary key index and for each entry in the index, it will follow the link to the data (the "bookmark") and find the flags and rowv, colv fields from the data page. This means a random disk access for each object in the photoobj table. Naturally, this will be excruciatingly slow since random access is several times slower than sequential access in disks.
If instead the optimizer picked a sequential scan of the whole photoobj table, the query could be completed within half an hour (assuming the server isnt badly loaded down). But with the chosen plan, it will take hours if not days! If you suspect that the optimizer is choosing the wrong plan, the way around it is to force the optimizer to ignore all the indices defined on that table. For example, you would rewrite the above query as follows:
Note that we made 2 changes to the original query:
5 hours vs 5 minutes!
Here is an example that demonstrates the dramatic speed improvement that this workaround provides. The following is a query submitted by Jon Loveday from the Astronomy Centre at the University of Sussex in the UK:
This query takes five and a half hours to run on the DR2 database, and involves a bookmark lookup on a phototag index. If you insert an (index=0) hint as shown below, it completes in 5 minutes!
Unfortunately, we are stuck with the bookmark bug for the time being. It will be fixed in the next version of SQL Server, SQL Server 2005 or Yukon, expected to be available in Fall 2004 for beta release. We expect the fix to be incorporated into the next SDSS Data Release (DR3) due to be made public in Jan 2005.
Roy Gal, Ani Thakar, Jim Gray, Alex Szalay
Last updated Apr 8, 2004.