Querying datasets

Datasets in a butler-managed data repository are identified by the combination of their dataset type and data ID within a collection. The Registry class’s query methods (queryDatasetTypes, queryCollections, queryDataIds, queryDatasets, and queryDimensionRecords) allow these to be specified either fully or partially in various ways.

Note

Registry queries utilize locally-cached information and heuristics to generate simpler queries and provide diagnostics when queries yield no results. Concurrent writes by other butler clients may not be reflected in these caches, if they happened since this Registry was initialized, and new datasets may not be found by queries as a result. Users can call Registry.refresh before querying to update the caches. Other Registry and Butler methods (Registry.findDataset and Butler.get variants in particular) do not suffer from this limitation; if caching is used in these contexts, we always fall back to database searches when cached information indicates that a dataset does not exist.

DatasetType expressions

Arguments that specify one or more dataset types can generally take any of the following:

• DatasetType instances;

• str values (corresponding to DatasetType.name);

• str values using glob wildcard syntax which will be converted to re.Pattern;

• re.Pattern values (matched to DatasetType.name strings, via fullmatch);

• iterables of any of the above;

• the special value “...”, which matches all dataset types.

Wildcards (re.Pattern and ...) are not allowed in certain contexts, such as Registry.queryDataIds and Registry.queryDimensionRecords, particularly when datasets are used only as a constraint on what is returned. Registry.queryDatasetTypes can be used to resolve patterns before calling these methods when desired. In these contexts, passing a dataset type or name that is not registered with the repository will result in MissingDatasetTypeError being raised, while contexts that do accept wildcards will typically ignore unregistered dataset types (for example, Registry.queryDatasets will return no datasets for these).

Collection expressions

Arguments that specify one or more collections are similar to those for dataset types; they can take:

• str values (the full collection name);

• str values using glob wildcard syntax which will be converted to re.Pattern;

• re.Pattern values (matched to the collection name, via fullmatch);

• iterables of any of the above, empty collection cannot match anything, methods always return an empty result set in this case;

• the special value “...”, which matches all collections;

Collection expressions are processed by the CollectionWildcard class. User code will rarely need to interact with these directly, but they can be passed to Registry instead of the expression objects themselves, and hence may be useful as a way to transform an expression that may include single-pass iterators into an equivalent form that can be reused.

Ordered collection searches

An ordered collection expression is required in contexts where we want to search collections only until a dataset with a particular dataset type and data ID is found. These include all direct Butler operations, the definitions of CHAINED collections, Registry.findDataset, and the findFirst=True mode of Registry.queryDatasets. In these contexts, regular expressions and “...” are not allowed for collection names, because they make it impossible to unambiguously define the order in which to search.

The CollectionWildcard.require_ordered method can be used to verify that a collection expression resolves to an ordered sequence.

Dimension expressions

Constraints on the data IDs returned by a query can take two forms:

• an explicit data ID value can be provided (as a dict or DataCoordinate instance) to directly constrain the dimensions in the data ID and indirectly constrain any related dimensions (see Dimensions Overview);

• a string expression resembling a SQL WHERE clause can be provided to constrain dimension values in a much more general way.

In most cases, the two can be provided together, requiring that returned data IDs match both constraints. The rest of this section describes the latter in detail.

The language grammar is defined in the exprParser.parserYacc module, which is responsible for transforming a string with the user expression into a syntax tree with nodes represented by various classes defined in the exprParser.exprTree module. Modules in the exprParser package are considered butler/registry implementation details and are not exposed at the butler package level.

The grammar is based on standard SQL; it is a subset of SQL expression language that can appear in WHERE clause of standard SELECT statement with some extensions, such as range support for the IN operator and time literals.

Expression structure

The expression is passed as a string via the where arguments of queryDataIds and queryDatasets. The string contains a single boolean expression which evaluates to true or false (if it is a valid expression). Expression can contain a bunch of standard logical operators, comparisons, literals, and identifiers which are references to registry objects.

A few words in expression grammar are reserved: AND, OR, NOT, IN, and OVERLAPS. Reserved words are not case sensitive and can appear in either upper or lower case, or a mixture of both.

Language operator precedence rules are the same as for the other languages like C++ or Python. When in doubt use grouping operators (parentheses) for sub-expressions.

General note — the parser itself does not evaluate any expressions even if they consist of literals only, all evaluation happens in the SQL engine when registry runs the resulting SQL query.

Following sections describe each of the parts in detail.

Literals

The language supports these types of literals:

Strings

This is just a sequence of characters enclosed in single quotation marks. The parser itself fully supports Unicode, but some tools such as database drivers may have limited support for it, depending on environment or encoding chosen.

Numbers

Integer numbers are series of decimal numbers optionally preceded by minus sign. Parser does not support octal/hexadecimal numbers. Floating point numbers use standard notation with decimal point and/or exponent. For numbers parser passes a string representation of a number to downstream registry code to avoid possible rounding issues.

Time literals

Timestamps in a query are defined using special syntax which consists of a capital letter “T” followed by quoted string: T'time-string'. Time string contains time information together with optional time format and time scale. For detailed description of supported time specification check section Time literals.

Range literals

This sort of literal is allowed inside IN expressions only. It consists of two integer literals separated by double dots and optionally followed by a colon and one more integer literal. Two integers define start and stop values for the range; both are inclusive values. The optional third integer defines stride value, which defaults to 1; it cannot be negative. Ranges are equivalent to a sequence of integers (but not to intervals of floats).

Examples of range literals:

• 1..5 – equivalent to 1,2,3,4,5

• 1..10:3 – equivalent to 1,4,7,10

• -10..-1:2 – equivalent to -10,-8,-6,-4,-2

Identifiers

Identifiers represent values external to a parser, such as values stored in a database. The parser itself cannot define identifiers or their values; it is the responsibility of translation layer (registry) to map identifiers into something sensible. Like in most programming languages, an identifier starts with a letter or underscore followed by zero or more letters, underscores, or digits. Parser also supports dotted identifiers consisting of two simple identifiers separated by a dot. Identifiers are case-sensitive on parser side but individual database back-ends may have special rules about case sensitivity.

In current implementation simple identifiers are used by registry to represent dimensions, e.g. visit identifier is used to represent a value of visit dimension in registry database. Dotted identifiers are mapped to tables and columns in registry database, e.g. detector.raft can be used for accessing raft name (obviously dotted names need knowledge of database schema and how SQL query is built). A simple identifier with a name ingest_date is used to reference dataset ingest time, which can be used to filter query results based on that property of datasets.

Registry methods accepting user expressions also accept a bind parameter, which is a mapping from identifier name to its corresponding value. Identifiers appearing in user expressions will be replaced with the corresponding value from this mapping. Using the bind parameter is encouraged when possible to simplify rendering of the query strings. A partial example of comparing two approaches, without and with bind:

instrument_name = "LSST"
visit_id = 12345

# Direct rendering of query not using bind
result = registry.queryDatasets(
    ...,
    where=f"instrument = '{instrument_name}' AND visit = {visit_id}",
)

# Same functionality using bind parameter
result = registry.queryDatasets(
    ...,
    where="instrument = instrument_name AND visit = visit_id",
    bind={"instrument_name": instrument_name, "visit_id": visit_id},
)

Types of values provided in a bind mapping must correspond to the expected type of the expression, which is usually a scalar type, one of int, float, str, etc. There is one context where a bound value can specify a list, tuple or set of values: an identifier appearing in the right-hand side of IN operator. Note that parentheses after IN are still required when identifier is bound to a list or a tuple. An example of this feature:

instrument_name = "LSST"
visit_ids = (12345, 12346, 12350)
result = registry.queryDatasets(
    ...,
    where="instrument = instrument_name AND visit IN (visit_ids)",
    bind={"instrument_name": instrument_name, "visit_ids": visit_ids},
)

Unary arithmetic operators

Two unary operators + (plus) and - (minus) can be used in the expressions in front of (numeric) literals, identifiers, or other expressions which should evaluate to a numeric value.

Binary arithmetic operators

Language supports five arithmetic operators: + (add), - (subtract), * (multiply), / (divide), and % (modulo). Usual precedence rules apply to these operators. Operands for them can be anything that evaluates to a numeric value.

Comparison operators

Language supports set of regular comparison operators: =, !=, <, <=, >, >=. This can be used on operands that evaluate to a numeric values or timestamps, for (in)equality operators operands can also be boolean expressions.

Note

The equality comparison operator is a single = like in SQL, not double == like in Python or C++.

IN operator

The IN operator (and NOT IN) are an expanded version of a regular SQL IN operator. Its general syntax looks like:

<expression> IN ( <item1>[, <item2>, ... ])
<expression> NOT IN ( <item1>[, <item2>, ... ])

where each item in the right hand side list is one of the supported literals or identifiers. Unlike regular SQL IN operator the list cannot contain expressions, only literals or identifiers. The extension to regular SQL IN is that literals can be range literals as defined above. The query language allows mixing of different types of literals and ranges but it may not make sense to mix them when expressions is translated to SQL.

Regular use of IN operator is for checking whether an integer number is in set of numbers. For that case the list on right side can be a mixture of integer literals, identifiers that represent integers, and range literals.

For an example of this type of usage, these two expressions are equivalent:

visit IN (100, 110, 130..145:5)
visit in (100, 110, 130, 135, 140, 145)

as are these:

visit NOT IN (100, 110, 130..145:5)
visit Not In (100, 110, 130, 135, 140, 145)

Another usage of IN operator is for checking whether a timestamp or a time range is contained wholly in other time range. Time range in this case can be specified as a tuple of two time literals or identifers each representing a timestamp, or as a single identifier representing a time range. In case a single identifier appears on the right side of IN it has to be enclosed in parentheses.

Here are few examples for checking containment in a time range:

-- using literals for both timestamp and time range
T'2020-01-01' IN (T'2019-01-01', T'2020-01-01')
(T'2020-01-01', T'2020-02-01') NOT IN (T'2019-01-01', T'2020-01-01')

-- using identifiers for each timestamp in a time range
T'2020-01-01' IN (interval.begin, interval.end)
T'2020-01-01' NOT IN (interval_id)

-- identifier on left side can represent either a timestamp or time range
timestamp_id IN (interval.begin, interval.end)
range_id NOT IN (interval_id)

The same IN operator can be used for checking containment of a point or region inside other region. Presently there are no special literal type for regions, so this can only be done with regions represented by identifiers. Few examples of region containment:

POINT(ra, dec) IN (region1)
region2 NOT IN (region1)

OVERLAPS operator

The OVERLAPS operator checks for overlapping time ranges or regions, its arguments have to have consistent types. Like with IN operator time ranges can be represented with a tuple of two timestamps (literals or identifiers) or with a single identifier. Regions can only be used as identifiers. OVERLAPS syntax is similar to IN but it does not require parentheses on right hand side when there is a single identifier representing a time range or a region.

Few examples of the syntax:

(T'2020-01-01', T'2022-01-01') OVERLAPS (T'2019-01-01', T'2021-01-01')
(interval.begin, interval.end) OVERLAPS interval_2
interval_1 OVERLAPS interval_2

NOT (region_1 OVERLAPS region_2)

Boolean operators

NOT is the standard unary boolean negation operator.

AND and OR are binary logical and/or operators.

All boolean operators can work on expressions which return boolean values.

Grouping operator

Parentheses should be used to change evaluation order (precedence) of sub-expressions in the full expression.

Function call

Function call syntax is similar to other languages, expression for call consists of an identifier followed by zero or more comma-separated arguments enclosed in parentheses (e.g. func(1, 2, 3)). An argument to a function can be any expression.

Presently there only one construct that uses this syntax, POINT(ra, dec) is function which declares (or returns) sky coordinates similarly to ADQL syntax. Name of the POINT function is not case-sensitive.

Time literals

Timestamps in a query language are specified using syntax T'time-string'. The content of the time-string specifies a time point in one of the supported time formats. For internal time representation Registry uses astropy.time.Time class and parser converts time string into an instance of that class. For string-based time formats such as ISO the conversion of a time string to an object is done by the Time constructor. The syntax of the string could be anything that is supported by astropy, for details see astropy.time reference. For numeric time formats such as MJD the parser converts string to a floating point number and passes that number to Time constructor.

Parser guesses time format from the content of the time string:

• If time string is a floating point number then parser assumes that time is in “mjd” format.

• If string matches ISO format then parser assumes “iso” or “isot” format depending on presence of “T” separator in a string.

• If string starts with “+” sign followed by ISO string then parser assumes “fits” format.

• If string matches year:day:time format then “yday” is used.

The format can be specified explicitly by prefixing time string with a format name and slash, e.g. T'mjd/58938.515'. Any of the formats supported by astropy can be specified explicitly.

Time scale that parser passes to Time constructor depends on time format, by default parser uses:

• “utc” scale for “iso”, “isot”, “fits”, “yday”, and “unix” formats,

• “tt” scale for “cxcsec” format,

• “tai” scale for anything else.

Default scale can be overridden by adding a suffix to time string consisting of a slash and time scale name, e.g. T'58938.515/tai'. Any combination of explicit time format and time scale can be given at the same time, e.g. T'58938.515', T'mjd/58938.515', T'58938.515/tai', and T'mjd/58938.515/tai' all mean the same thing.

Note that astropy.time.Time class imposes few restrictions on the format of the string that it accepts for iso/isot/fits/yday formats, in particular:

• time zone specification is not supported

• hour-only time is not supported, at least minutes have to be specified for time (but time can be omitted entirely)

Examples

Few examples of valid expressions using some of the constructs:

visit > 100 AND visit < 200

visit IN (100..200) AND tract = 500

visit IN (100..200) AND visit NOT IN (159, 191) AND band = 'i'

(visit = 100 OR visit = 101) AND exposure % 2 = 1

visit.datetime_begin > T'2020-03-30 12:20:33'

exposure.datetime_begin > T'58938.515'

visit.datetime_end < T'mjd/58938.515/tai'

ingest_date < T'2020-11-06 21:10:00'

Query result ordering

Few query methods (queryDataIds and queryDimensionRecords) support special constructs for ordering and limiting the number of the returned records. These methods return iterable objects which have order_by() and limit() methods. Methods modify the iterable object and should be used before iterating over resulting records, for convenience the methods can be chained, see example below.

The order_by() method accepts a variable number of positional arguments specifying columns/fields used for ordering, each argument can have one of the supported formats:

• A dimension name, corresponding to the value of the dimension primary key, e.g. "visit"

• A dimension name and a field name separated bey a dot. Field name can refer to any of the dimension’s metadata or key, e.g. "visit.name", "detector.raft". Special field names "timespan.begin" and "timespan.end" can be used for temporal dimensions (visit and exposure).

• A field name without dimension name, in that case field is searched in all dimensions used by the query, and it has to be unique. E.g. "cell_x" means the same as "patch.cell_x".

• To reverse ordering for the field it is prefixed with a minus sign, e.g. "-visit.timespan.begin".

The limit() method accepts two positional integer arguments - limit for the number of returned records and offset (number of records to skip). The offset argument is optional, if not provided it is equivalent to offset 0.

Example of use of these two methods:

# Print ten latest visit records in reverse time order
for record in registry.queryDimensionRecords("visit").order_by("-timespan.begin").limit(10):
    print(record)

Error handling with Registry methods

Registry methods typically raise exceptions when they detect problems with input parameters. Documentation for these methods describes a set of exception classes and conditions in which exceptions are generated. In most cases, these exceptions belong to one of the special exception classes defined in lsst.daf.butler.registry module, e.g. DataIdError, which have RegistryError as a common base class. These exception classes are not exposed by the lsst.daf.butler module interface; to use these classes they need to be imported explicitly, e.g.:

from lsst.daf.butler.registry import DataIdError, UserExpressionError

While class documentation should list most commonly produced exceptions, there may be other exceptions raised by its methods. Code that needs to handle all types of exceptions generated by Registry methods should be prepared to handle other types of exceptions as well.

A few of the Registry query methods (queryDataIds, queryDatasets, and queryDimensionRecords) return result objects. These objects are iterables of the corresponding record types and typically they represent a non-empty result set. In some cases these methods can return empty results without generating an exception, for example due to a combination of constraints excluding all existing records. Result classes implement explain_no_results() method which can be used to try to identify the reason for an empty result. It returns a list of strings, with each string a human-readable message describing the reason for an empty result. This method does not always work reliably and can return an empty list even when result is empty. In particular it cannot analyze user expression and identify which part of that expression is responsible for an empty result.