Certify Constraints

Certify Overview

This section documents the design and syntax of CRDS certify .tpn constraint definition files. The .tpn constraints define many of the checks CRDS applies to rule or reference files. CRDS certify performs these forms of checks:

  1. .tpn defined constraint checks unique to CRDS

  2. CRDS type spec defined unique table row checks, missing mode checks

  3. fitsverify checks, recategorized as needed

  4. FITS format checks

  5. .rmap update checks

  6. sha1sum checks (bit-for-bit identical file rejection)

This section discusses syntax for (1), file constraints CRDS defines itself internally.

Overview of .tpn Files

tpn files are used by CRDS certify to define checks on reference files and rmaps. For HST .tpn files define almost all CRDS checks (some table checks also exist) and were incorporated verbatim from CDBS certify. This is the origin of CRDS .tpn syntax.

.tpn File Organizations

For HST, .tpn files were written for every combination of <instrument>_<type>.

There are two forms of .tpn, one which constrains reference file properties (.tpn) and one which constrains rmap properties (_ld.tpn). The _ld.tpn files originally constrained CDBS database field values for matching parameters; because of translations between FITS headers and database values, reference file and database values were not necessarily identical. For JWST and Roman, .tpn and _ld.tpn files are almost universally identical.

While HST’s design of limiting the file structure to <instrument>_<type> combinations was simple, there was also a significant penalty of redundant definitions for keyword constraints like &USEAFTER or &PEDIGREE.

.tpn Directives

Explicit Directives

Typical constraint directives result in a CRDS TpnInfo() object being defined in CRDS certify, which corresponds 1:1 to a CRDS Validator() object/subclass. The TpnInfo() is a bundle of constraint information, a Validator() is a class which checks the constraint defined by a TpnInfo. Ultimately the outcome of .tpn files is a list of TpnInfo() constraints to check, which are tuples of properties, interpreted via appropriate Validator() subclassses.

Synthetic Directives

The CRDS certifier uses the JWST/Roman data models in two ways:

1. CRDS certify calls datamodels.open() on each reference to directly apply data models checks. This is very simple and fairly robust, but suffers from the dependency of open() on the DATAMODL keyword which is used to specify the model class of the reference file. The model class implies the exact schema to check. If DATAMODL is not specified, open() won’t validate specialized schema.

2. Second, CRDS loads the data models core schema and converts them into synthetic TpnInfo() objects which don’t appear in a .tpn file. These synthetic constraints are interpreted by CRDS normally to make that checking visible and also to permit automatically promoting the constraints from “optional” (check if present) to “definintely required” where keywords are needed for CRDS matching. The synthetic TpnInfo() objects generally scrape value enumerations from the core schema for use in CRDS.

Include Directive

The semantics of ‘include’ are roughly modeled after the C pre-processor’s #include directives.

The include directive permits one .tpn file to include the text of another as if the directive was replaced by the contents of the included file. This permits factoring out common constraints into a single file and using them in multiple locations, e.g.:

include nir_sci_array.tpn

defines constraints on the SCI array for all NIR instruments except NIRSPEC.

For the sake of simplicity, the line of the directive should begin with ‘include’ (unquoted), a single space, and the name of the included .tpn file (also unquoted). Like many CRDS directives it is space delimited and consists of exactly two words, written on a single line.

An included file should also be located in the .tpn includes directory and generally will not follow the JWST classes of include files (all_all, fgs_all, etc.) or their naming conventions.

Replace Directive

The replace directive causes CRDS to do substitutions on all subsequent lines in the .tpn file, even lines included from other .tpn files. The replace directive consists of 3 words in the general form of:

replace <original_pattern> <replacement>

where an example directive is:

replace SCI COEFFS

which means that all subsequent instances of the pattern SCI in a particular .tpn (or included .tpns) should be replaced by COEFFS. In this case, it effectively rewrites constraints on the SCI array as constraints on the COEFFS array.

An important limitation of ‘replace’ to note is that it only applies within the textual extent of on particular file class. At this time, it’s not possible to e.g. define a replacement in all_all and have it apply everywhere.

Constraint Directive

By far the most common directive is the constraint directive, which defines one condition on a reference file and is of the form:

<name>  <keytype>  <datatype>  <presence>   <value constraint>

or sometimes with the values omitted:

<name>  <keytype>  <datatype>  <presence>

Before explaining each field in detail, a typical (but abbreviated) example constraint taken from JWST is:


The columns of the constraint are interpreted as follows:

  1. The first space delimited word defines the keyword / data model path: META.INSTRUMENT.PUPIL

  2. The second word defines this as a (H)eader keyword constraint vs. e.g. (C)olumn

  3. The third word defines this as a (C)haracter keyword value vs. e.g. (F)loat

  4. The fourth word defines this as (O)ptional, it may be omitted. Another common value is (R)equired. An expression may also be used in this slot to define if/if-not the constraint should apply at all; additional semantic refinements may also be added by wrapping helper functions.

  5. The final “word” is a comma separated list of values. Multiple lines may be used by terminating each line with a backslash except the final line. In some cases the value list is replaced by a Python expression which should evaluate to True. Literal numerical ranges may also be specified.

Whitespace in Constraint Fields

Since whitespace is used to delimit fields within a constraint, spaces / whitespace MAY NOT appear within any single field, i.e. the value list, presence expressions, constraint expressions, etc. This can be slightly awkward at times but the addition of extra parentheses to existing punctuation is generally sufficient to write expressions containing no spaces.

For example, even expressions such as:

("IFU" not in EXP_TYPE)

which contain instrinsic whitespace can often be rephrased in a workable way as:


A limitation of this approach is that literal strings containing white space are not permitted/straightforward in expressions. In that area, writing additional helper functions or custom validators might provide a way out.

While the idea of modernizing .tpn syntax is pretty obvious, the downsides of switching to more readable file formats like JSON or YAML are a combination of:

* Up-front work
* Additional testing for multiple projects
* Constraints which become more verbose and less dense.

Since constraints are easier to read and consider en masse when they’re expressed in a dense format, not even readability is a complete slam dunk as a motivation for modernizing formats.

<Name> Field

The name field specifies the name of a header keyword, table column, array, or expression constraint.

Header Keyword Names

Header and table keyword names correspond roughly to FITS keywords or JWST CAL data models paths flattened into a single string, e.g.:

Array Names

Array names are specified as the bare HDU name in the <name> field, e.g. SCI. These are referenced within expressions as <name>_ARRAY. These are case insensitive and specified in all capital letters, numbers, or underscores much like FITS keywords. They should begin with a letter and be valid program identifiers.

There are two additional specification cases for array names:

  1. FITS extensions can also named by number, e.g. EXT1 or EXTENSION1 refers to the data of HDU #1. In constraint expressions these are referred to as e.g. EXT1_ARRAY. These can be discriminated from normal header keywords by the keytype, which will be array vs. header.

  2. FITS extensions can be named by (name, ver), in CRDS this is denoted as <name>__<ver>, which corresponds to e.g. (‘SCI’, 1). In constraint expressions they are referred to as as <name>__<ver>_ARRAY, e.g. SCI__1_ARRAY. These can be differentiated from normal array extension names by the double-underscore-digit convention, an imperfect compromise.

Expression Constraint Names

Expression constraint names describe the check performed by the value expression, they do not describe any physical entity within the reference file. Note that expression here refers to a keytype=X constraint and have no relation to expressions used in the <presence> field described below.

<Keytype> Field

The keytype field consists of one character corresponding to:

keytypes = {
   "H" : "HEADER",
   "C" : "COLUMN",
   "G" : "GROUP",
   "A" : "ARRAY_FORMAT",
   "D" : "ARRAY_DATA",
   "X" : "EXPRESSION",
Header Keywords (H)

Header keyword names correspond to values taken from the union of all HDU headers.

Almost all of the HST constraints taken from CDBS are enumerations applying to a single FITS or GEIS keyword, e.g. READPATT.

Many JWST and Roman constraints are written using the format independent (FITS, ASDF, JSON…) data model hiearchical path names munged for CRDS purposes into all capital letters with periods replaced by underscores so that they can be evaluated as a single keyword name rather than as nested objects.

Columns (C)

Column names generally apply to the name of a FITS table column and the corresponding constraint applies only to the values of that single column in isolation.

Column expressions

Array Format (A)

Array format constraints apply to lightweight array properties taken from FITS HDU data:

    SHAPE = hdu.data.shape,
    KIND = generic_class,
    DATA_TYPE = typespec,
    COLUMN_NAMES = column_names,
    EXTENSION = i,
    DATA = None,

Most notably, the array data itself is not available for constraint checking but the lightweight properties are relatively fast and small to load.

Generally, array format and data keytypes have expression constraints rather than value enumerations, ranges, etc. Most commonly expressions limit the array shape and type.

Array expressions can be written in terms of all arrays for which constraints are defined. So an ERR array constraint might also refer to SCI if it was known to be loaded elsewhere.

Array Data (D)

Array data checks are heavy weight and entail loading the actual reference data so that constraints can be applied to it:

    SHAPE = hdu.data.shape,
    KIND = generic_class,
    DATA_TYPE = typespec,
    COLUMN_NAMES = column_names,
    EXTENSION = i,
    DATA = hdu.data      #  XXX the difference between 'A' and 'D' constraints!

Generally, array format and data keytypes have expression constraints rather than value enumerations, ranges, etc. Most commonly expressions limit the array shape and type.

Expressions (X)

Expressions replace the typical value enumeration, range, etc. with a Python expression written in terms of the reference file header and array properties. While A and D array constraints are also generally written as as expressions, in contrast, an X constraint loads no new array properties and includes no arrays. The value expression should be written in terms of header keywords only. Arrays are pre-loaded and remain available to all expressions for the duration of a single reference file check.

Group (G)

Not implemented but parsed for the sake of HST CDBS backward compatibility.

<Datatype> Field

The datatype field conceptually corresponds to the type of a FITS/ASDF keyword defined in the reference file header or table. Similar properties are imposed on data models paths/keywords which may or may not correspond to a FITS/ASDF keyword.

The datatype is written as a single character with these translations:

datatypes = {
   "C" : "CHARACTER",      #  ignores case differences
   "I" : "INTEGER",
   "L" : "LOGICAL",
   "R" : "REAL",           #  float32 value(s)
   "D" : "DOUBLE",         #  float64 value(s)
   "X" : "EXPRESSION",     #  constraint expression expected

The X datatype indicates that the constraint will be a boolean expression and hence has no data type. For ‘A’ and ‘D’ keytypes, the expression is abstract, referring to no particular keyword or array by definition. In the case of ‘C’ keytypes, the expression is only permitted to refer to the current column value.

<Presence> Field

The presence field determines the conditions under which a constraint applies and what should happen when it is omitted:

presences = {
    "E" : "EXCLUDED",
    "R" : "REQUIRED",
    "P" : "REQUIRED",
    "W" : "WARN",
    "O" : "OPTIONAL",
    "F" : "IF_FULL_FRAME",
    "S" : "IF_SUBARRAY",
    "A" : "ANY_SUBARRAY"
Simple Presence Values

Simple presence values are specified as a single character which correspond to these classifications:

REQUIRED or True results in an error if the keyword is not present in the file header or tables or is UNDEFINED or the constraint is not satisfied.

False means a constraint does not apply.

WARN results in a warning if the keyword is not present or is UNDEFINED.

OPTIONAL indicates that a constraint should be satisfied if the keyword is present and not UNDEFINED but is not an error when omitted.

IF_FULL_FRAME means that the constraint only applies when SUBARRAY keywords are defined (SUBARRAY,SUBSTRT1,SUBSTRT2,SUBSIZE1,SUBSIZE2) and SUBARRAY describes a full frame (FULL,GENERIC,N/A,ANY,*).

IF_SUBARRAY means that the constraint only applies when SUBARRAY keywords are defined and SUBARRAY does not describe a full frame.

ANY_SUBARRAY means that the constraint only applies when SUBARRAY keywords are defined.

EXCLUDED means that a keyword should not be specified and was supplied for backwards compatibility with HST CDBS and is generally unused.

For HST, every instrument and type specified the presence requirement for every keyword. This resulted in value enumerations repeated over and over throughout the .tpn files.

For JWST and Roman, CRDS support specifying keywords as optional… with one twist: if an optional keyword is used by an rmap to perform matching (appears in the ‘parkey’ header field), then every optional constraint on that keyword for that particular reftype becomes required.

This permits constraints to be specified once as optional at a relatively global level for easier maintenance, but then become “required” if a particular reftype uses the keyword directly within CRDS for matching. (This is a reflection of the “prime directive” of the CRDS certifier: while general checks can be implemented, the most crucial aspect of CRDS checking is to ensure that files work within CRDS. Although CRDS does strive to implement additional checks, the only real measure that references will work with the CAL code is running calibrations.)

For even more control, or for keywords not used by CRDS matching, additional constraints can be defined in more specialized .tpn’s.

Presence Expressions and Helpers

A Python expression can be specified to define when a constraint does or doesn’t apply based on keyword values.

The expression should begin with ( and end with ) and should contain no spaces. (Sometimes extra parens are required to break up the expression into words using punctuation instead of spaces.)

An example of a presence expression is:


which means that the constraint only applies when EXP_TYPE is not FGS_ID-STACK.

Keyword names used in presence expressions follow the usual rules and must be valid Python identifiers in all caps. Periods from data model paths are replaced by underscores to make the paths into simple identifiers suitable for Python’s eval().

Presence helpers have been defined to convert the boolean result of a presence expression into a simple presence value. This enables conditional optional keywords, conditional warnings, conditional subarray expressions, etc:

optional(expr)     -->   False or 'O'
full_frame(expr)   -->   False or 'F'
subarray(expr)     -->   False or 'S'
any_subarray(expr) -->   False or 'A'
required(expr)     -->   False or 'R'
warning(expr)         -->   False or 'W'

For example, an expression further refined by the full_frame() helper:


means that fundamentally, it only implies when EXP_TYPE is not FGS_ID-STACK, but in addition, it only applies when SUBARRAY keywords are defined and the SUBARRAY is some form of full frame, e.g.FULL or GENERIC. In effect, the helper arranges things so that the presence field is ‘F’ if the wrapped expression is satisfied.

Note that an expression return value of False indicates a constraint does not apply at all. An expression return value of True indicates the constraint is REQUIRED.

Helper functions in .tpn files are distinguished by being written in all lower case; this prevents collisions with keyword, column, or array names which are always written in upper case.


The <values> field of each constraint can define a number of things, including enumerations of literal values:


numerical ranges:


constraint expressions:


custom validator identifiers:


or nothing at all.


Value enumerations list the possible literal values that can be assigned to a keyword, e.g.:


Ranges specify inclusive numerical ranges which keyword values must lie within, e.g.:


means the value should be within 1 and 10 inclusive. An equivalent expression constraint would be:


where KEYWORD is the name of the constrained keyword.

Custom Constraint Validators

Custom constraint handlers define new classes of validators and are always specified by a value / validator name beginning with &, e.g.:

META.USEAFTER   H   C   R               &JWSTDATE

where validator values have meanings like:

&PEDIGREE  -- implements algorithm to check various PEDIGREE value forms
&USEAFTER  -- implements HST USEAFTER date/time format checking
&JWSTDATE  -- implements JWST/Roman datetime format checking,  e.g. JWST USEAFTER

NOTE: Roman also uses the &JWSTDATE format validator.

Custom constraint validators can perform arbitrary processing to validate a single keyword value, i.e. specify precise date formats, etc. Custom constraint validators are defined in the crds.certify.validators submodules with classes named like e.g. PedigreeValidator, UseafterValidator, JwstdateValidator.

Expressions Constraints

Unlike presence expressions which define when a constraint should or should not be applied, expressions constraints define the condition which should be satisfied when the constraint is applicable.

Someone might briefly wonder if both presence and constraint value expressions are needed. The answer is “yes” because a negative result of a value expression is limited to “constraint failed” while a negative result for the presence expression is limited to “do not evaluate”, so the concerns truly are separate and two expressions are needed.

Constraint expressions always begin with ‘(’ and end with ‘)’ and should contain no spaces.

An example expression constraint is:


which asserts that XSTART + XSIZE - 1 should fall within the boundaries of the detector’s 2048 X-dimension.

When specified within CRDS .tpn files, JWST CAL data models paths (ie. keyword names) are flattened to simple strings that resemble FITS keywords in all upper case:

meta.subarray.xstart -->  META.SUBARRAY.XSTART

Within expressions, the periods are replaced with underscores:


so that when the name is eval()’ed it is a simple Python identifier instead of a e.g. three nested objects.

Array identifiers appear in expression constraints as e.g. SCI_ARRAY to refer to the SCI HDU properties. In this case SCI_ARRAY is a true utils.Struct() object so it refers to Struct() properties within the eval() expression using normal Python object attribute access, e.g. SCI_ARRAY.SHAPE not SCI_ARRAY_SHAPE.

In expression constraints over (C)olumn keytypes, the only variable available is VALUE, which contains the column value currently under consideration.

Expression warn_only() Mutator/Wrapper

Expression constraints have the unique property that they can be mutated to generate warnings instead of errors. In contrast, there is no such mechanism for value enumerations or ranges. Custom validators can be written to issue warnings as needed.

The example constraint expression above can be mutated to a warning like this:


If the constraint fails, a log ERROR which would fail the file submission is replaced with a WARNING which can be investigated and/or ignored. Warnings truly are warnings, they can flag fatal conditions but may not be applicable in all cases.

Note that this is distinct from the Presence field “W” designator and related warning() mutator which only alter the “required” status of a constraint/keyword, not the result of a constraint failure.

Table Expression Helpers

Expression helper functions were added to check basic table properties based on the contents of HDUS. To some degree these are redundant to the HST “C” column style constraints… but have the advantage that they operate directly on HDU array properties and type information. In contrast, the “C” column constraints followed HST practices relying more on value string formatting, e.g. “if it looks like a FLOAT, it is a FLOAT.” In practice, file developers actually do make the error of adding FLOAT repr()’s to references instead of actual FLOAT values so this minor extension was added to enable checking that.

Some of the table helpers:


Empty Value Lists

The value list can be empty, in which case the constraint is limited to checking presence and type.

Unique Row Table Checks

CRDS has an HST requirement to attempt to detect missing table modes. This is done by specifying table columns which should identify unique rows, and then comparing the unique rows of and old and new table to see if any unique rows are dropped. The same generic capability can also be used by JWST.

Because the table row checks are crude approximations, the net result is generally one of two kinds of warnings. First, a table may define more than once instance of a row which should be uniquely identified; these are referred to in a warning as “duplicate” rows. Second, the new version of a table may drop unique rows found in the original version; this is reported loosely as one or more missing “modes”.

Unique rows are defined by combinations of column parameters. The parameter names used to select unique rows are defined in the “spec” file of each reference type as needed in a unique_rows header field, e.g.:

miri_cubepar.rmap:    'unique_rowkeys' : ('GRATING', 'FILTER'),

The spec files are located in the “specs” directory of each project directory, e.g. crds/jwst/specs/miri_cubepar.rmap. Spec files also define other static reference type properties like short and long form names, etc. To speed loading on slow file systems, specs for all instruments and types are combined into a single combined_specs.json file for each project.

Because one reference may define more than one table, unique row names are only used in the row selection combination if they’re present in a particular table. Independent .tpn checks can verfy that all required columns are present.

In the above example, if one table defined unique rows by GRATING, and a second table defined unique rows by FILTER, CRDS would correctly support both table checks. In a different situation, unique table rows might be defined by combinations of both FILTER and GRATING. This dicey interpretation of unique rows turns out to be good enough in practice, it’s relatively uncommon to check multiple tables in one reference.

Debugging Certify Updates

When run without –verbose, CRDS certify is relatively quiet about what it is checking unless checks fail. (A current exception which may change is the regurgitation of the complete fitsverify output. But most .tpn checks are silent unless –verbose is set or they fail.)

Verifying changes to CRDS certify .tpn files can generally done by running certify over some context, imap, or rmap in –deep mode which will attempt to certify each reference file and/or sub-mapping. Further, turning on the debug messages with –verbose or –verbosity=60 or 70 or.. will generate output on what CRDS is checking, how, and why / why not.

An example of running CRDS this way would be:

$ export CRDS_SERVER_URL=https://hst-crds.stsci.edu
$ export CRDS_PATH=/grp/crds/cache
$ crds certify hst-acs-biasfile-edit --deep --dump-unique-errors --verbose --dump-provenance

The output, which is copious, is relatively self-explanatory. Typically one greps through it for output from the constraint being added or modified.

For extensive changes to certify, it can be useful to run it on all the active reference files like this:

Where the symbolic context name ‘hst-edit’ is interpreted to something more literal like ‘hst_0442.pmap’:

$ export CRDS_SERVER_URL=https://hst-crds.stsci.edu
$ export CRDS_PATH=/grp/crds/cache
$ crds certify hst-edit --deep --dump-unique-errors --verbose --dump-provenance

Likewise, exhaustive testing may require running certify on ‘hst-edit’ as well after setting:

$ export CRDS_SERVER_URL=https://hst-crds.stsci.edu