Skip to content
Switch branches/tags
Go to file
Cannot retrieve contributors at this time
309 lines (239 sloc) 14.1 KB

Tabular Overview

At the core, tabular knows nothing about rendering. The base-level library, in the repo's top-level, can be imported and used to create tables and add content. Rendering requires using another layer to convert the output for display.

A Table is an interface. There is one core public type which implements the interface. This allows the core type to be embedded in the wrapper/display objects and for those to satisfy the table interface, thus being tables themselves. Rows and cells are not interfaces. The core public type for a table is, imaginatively, *ATable.

Callers doing simple table usage should import the auto sub-package and use the RenderTable interface type, or use tabular's code Table interface (which is part of RenderTable).

A table consists of rows of cells and some metadata. The metadata includes virtual columns, allowing for addressing by column too. Columns are identified by the header name, or numerically starting at 1, with column 0 being reserved for use in some contexts to mean "applies to column, is default for all columns". There is only one (or zero) header row per table.

Errors in adding data are usually not reported immediately, to let data stream in. Instead, errors accumulate in an error holder. Rows hold errors, but once a row is part of a table, its errors become the tables' errors (and the error container is diverted to be the table's). An error container can be interrogated for its list of current errors.

The errors are either a list of non-nil errors, or nil. An empty list should never be returned. If a nil is returned in the list of errors then that is a bug in tabular.

A row either contains cells or is a "special" row. The only type of special row is a "separator" row. The tabular layer itself doesn't know what a separator row is, beyond that it exists and a row can be one. The AddSeparator() table method adds one, the IsSeparator() row method asks a row if it is one. The Cells() method, which returns an array of cells, can return nil if and only if the row is special (ie, at present, a separator). A real row is always a splice of cells, even if that splice is empty.

A Cell contains "an object". That object can be a string, something which satisfies Stringer or GoStringer, a rune, or another Cell. Cells can contain cells and this is intended to allow for dynamic update, based upon evaluation. Defining a MarshalText method may be advisable and a future API bump might choose to prefer MarshalText to Stringer or GoStringer.

If a Cell contains an object then various rendering layers may make use of other interfaces satisfied by that object to determine how to display it; loosely, think of "width" and "height", but this will be covered in more detail below.

Cells, Rows, Columns and Tables can have "properties" set upon them. Properties are namespaced objects, very similar to Golang's net contexts. Clients of the tabular package are free to decorate items with whatever properties they want.

The tabular package supports automatically updating properties at "addition" time and at "render" time. This is done by setting callbacks. Callbacks can be on a table or a row. Within the table, they can be registered for use on a table or a column, for when a row is added, or when a cell is added.

The cell's location in the grid is not a property, but is available via a method call upon the cell.

The callbacks and properties should not be exposed to end-users. This should be considered a low-level tinkering API and your glue layer should mask the details.

All child objects have links back to their containers. This is used, eg, to be able to get column information for a given cell. This does mean that there are ownership loops.

Sub-package Commonalities

In all cases:

  • There is a Wrap() function which takes any tabular.Table and returns a wrapper object for this sub-package.
  • There is a New() function which generates an empty table for this sub-package.
  • The wrapper objects have Render() and RenderTo() methods, and the packages have top-level functions which create a wrapper object, with default options, and calls the object methods.
  • The Render() method will return a string of the rendered text, together with an error.
  • The RenderTo(io.Writer) method uses a stream-based approach and only returns an error.
  • The wrapper object's type is named with a FooTable naming style, accepting that this causes some stuttering. This is acceptable because most callers should never need to specify the type, but instead be using the tabular.Table interface if they care at all beyond letting the type be inferred. With so many variants, I went for clarity over smoothness in reading out the fully-qualified type name.

Because the sub-package New() returns an object which satisfies the tabular.Table interface, it should be capable of being populated like any other, and most callers with simple use-cases should be able to only import the sub-package, not tabular itself.

It is possible to use table properties to store data, but that's more complexity than is usually warranted. If you want static attributes, put them in your wrapper object. If you want dynamic attributes, updated based upon content, then use properties, and consider how to hide this from your users.


This is the auto sub-package of tabular.

This provides a RenderTable interface, which adds Render() and RenderTo() methods to the core tabular.Table.

Unlike the more-specific sub-packages, the New() and Wrap() methods take a string argument. The string is a style. This string is taken to be a dot-joined sequence of sections, where the first section is a sub-package or a texttable/decoration.Decoration.

So auto.New("csv") returns an auto.RenderTable which is satisfied by a *csv.CSVTable. auto.New("texttable.utf8-light") is equivalent to auto.New("utf8-light") and returns an auto.RenderTable which is satisfied by a *texttable.TextTable with the decoration set to utf8-light.

auto.New("") is currently equivalent to auto.New("csv") but future extensions might pass the foo onto some appropriate initialization of the csv package.

The auto sub-package does not provide specific implementations of the Render or RenderTo sub-packages, but does provide the usual package-level wrappers.

In addition, auto.ListStyles() returns a sorted list of strings, each of which is a valid input for auto.New(style). The list is not guaranteed to be exhaustive, but should cover the common cases. It is guaranteed to be exhaustive of all top-level style names (the bit before the first .).

Text Table Display

This is the texttable sub-package of tabular.

This system is designed to draw a pretty table on a cell-based display system, such as a classic Unix terminal emulator. Every "display cell" (not table cell) is a fixed pixel width and height, so using box-drawing characters, everything can be made to line up.

If a cell's object supports the Height() method then that overrides a calculation based on "newlines count + 1". If a cell's object supports the TerminalCellWidth() method then that overrides a calculation based on text, figuring out the longest line (multi-line supported) where length is Unicode-aware and display-width (wide char and combining char) aware.

Calculations upon cells are done at render time, to examine the contents and determine width and height for text-table purposes. These are stored as properties of each cell. This is a complete table sweep before printing the first line starts. Then the rendering uses the properties to size itself and print the table.

TODO: The maximum widths should become column properties of the table and the height become a row property.

There is a decoration sub-package of texttable which has decoration styles for rendering tables, as ASCII or as a few varieties of Unicode box-drawing. Decoration objects can be created by callers and set directly upon the table, or can be set by name. The names are maintained as a registry within the decoration package. Each name is a simple string, thus typos are a potential source of errors. For the styles native to the decoration package, package constants are exported with the names, permitting compile-time checks to catch issues. Eg, use decoration.D_UTF8_LIGHT_CURVED instead of "utf8-light-curved" if you are willing to import the decoration package here. There's a trade-off between provable correctness and importing more and clients get to choose the level they're happy with.

HTML Table Rendering

This is the html sub-package of tabular.

It does not render "separator" rows.

Table Id, Class and Caption are top-level attributes.

The table is rendered using Golang's html/template to handle auto-escaping of unsafe data. If the template name matters (you are using templates more generally) then you can use TemplateName on the HTMLTable object.

There is an SetRowClassGenerator() method to let you register your own function to be used to emit a class-name on each <tr> of the table's body. See the package docs for more details (you get a row index and your own context for passing state).

CSV Rendering

This is the csv sub-package of tabular.

The rendering is compliant to RFC4180. All fields are always quoted. Note in particular that newlines within strings are not \n escaped and double-quotes are doubled for escaping, thus "". Both of these attributes are RFC-specified.

The CSVTable type is designed to be extensible to change separators, escaping styles and more. The default is RFC4180, and that's the only current style, but we should accept PRs for any sane options, and also "family" sets, as long as well-specified. At present, only the fieldSeparator is called out in the struct, and there are no mutators for it. That's not a bug, just "not yet implemented, waiting for solid use-cases".

JSON Rendering

This is the json sub-package of tabular.

This emits structured JSON where the table is represented as an array of objects, one object per row, and the string representations of the column headers become the object keys. It's an error to not have headers; it's an error to have missing, empty or duplicate headers (as rendered to string).

This rendered passes the underlying stored items within the cells to encoding/json for marshalling, so will handle arbitrary types; because we have only required that cell types support .String() string, we have a fallback of using String() if the marshalling returns a sequence of {} corresponding to an empty struct, ie a struct with no exported fields. If you store a struct with exported fields and have previously relied upon String() being defined, then the output in JSON format will be less than ideal unless you also define a MarshalText() method.

There is no handling for a MarshalText() or MarshalJSON() method choosing to return {}, on the assumption that if they do so, then a String() method would also choose to return {}. This is currently a brute fallback for a heuristic to Do The Right Thing, rather than reflection-aware handling.

A future revision might switch away from generic JSON marshalling and simplify this; if we do so, then we'll support the generic MarshalText() but not the MarshalJSON() legacy method.

Markdown Rendering

This is the markdown sub-package of tabular.

The dialect emitted is "GitHub Flavored Markdown" tables.

For automated production you should avoid this renderer in favor of the html renderer; the markdown table support is extremely limited and there is no definition of what should happen for a number of otherwise valid inputs.

The expectation is that documentation maintainers can use this package with their client tool to get output which can be inserted in markdown and be subject to human review. If you don't have a human in the loop, then there's absolutely no reason to not just use the HTML renderer: Markdown passes through HTML, so the HTML output is more portable and safer.

But if your client tooling supports auto and can just be asked "hey, give me the markdown" then documentation maintainers can use that for grabbing samples.

Core Properties

Some properties are available within the properties sub-package of tabular. These properties affect core rendering and are interpreted appropriately within tabular itself.

    • Skipability:
      • Skipable is the property key, value must be a boolean
      • non-boolean values may result in panic
      • currently only a column property, and is permitted to be used by renderers to control whether or not to skip a column entirely, or for the JSON rendered to skip including a field for items in that column if empty.
        • At present, is only used by JSON, have not yet implemented a multiple-pass system to let it be used for entire-column suppression.
    • controls for text alignment; currently this is limited to very simplistic Left, Center and Right values, which may be used as values for the align.PropertyType key. This is expected to change to become more flexible, but this simple use-case will be grandfathered in to remain simple.
    • This is currently only a column property.

Coding Style

  1. All code should pass go fmt and go vet
  2. golint is too opinionated in ways I care about and is explicitly not a fixed goal. Reducing its complaints for new code is worthwhile, unless that contradicts something here. Changing an API to match golint's style is never acceptable, unless there's an API major revision bump.
  3. Exported constants should be ALL_CAPS
  4. Imports should be in batched groups, so that go fmt will sort within each group but not move between groups; those groups should be:
  5. stdlib packages
  6. testing-related packages, if we're a test
  7. intra-repo/org packages
  8. external third-party packages