Developer experiences from the trenches
Fri 09 May 2025 by Michael Labbe
tags code
Every modern game is powered by structured data: characters, abilities, items, systems, and live events. But in most studios, this data is fractured—spread across engine code, backend services, databases, and tools, with no canonical definition. Your investment in long-lived game data is tied to code that rots.
Grain DDL fixes that. Define your data once and generate exactly what you need across all systems. No drift. No duplication. No boilerplate. Just one source of truth.
Game data is what gives your game its trademark identity — its feel, the sublety of balance, inventory interactions, motion, lighting and timing. Game data is where your creative investment lives, and you need to plan for it to outlive the code.
Grain DDL is a new data definition language — a central place to define and manage your game’s data models. Protect your investment in your game data by representing it in a central format that you process, annotate, and control.
Typically, games represent data by conflating a number of things:
Game data has a lifecycle: represented by a database, sent across a network wire, manipulated by a schema-specific UI and serialized to disk. In each step, a developer binds their data schema to their implementation.
In each of these steps, there is imperative code that binds the schema to an implementation. More importantly, there is no longer a canonical representation of a central data model. There are only partial representations of it strewn across subsystems, usually maintained by separate members of a game team.
Grain DDL protects your data investment by hoisting the data and its schema outside of implementation-specific code. You write your code and data in Grain DDL, and then generate exactly the representation you need.
Grain DDL comes in two parts: a data definition language, and Code Generation Templates. Consider this simple Grain definition:
struct Person {
str name,
u16 age,
}
From there, you can generate a C struct with the same name using code generation templates:
/* for each struct */
range .structs
`typedef struct {`.
/* for each struct field */
range .fields
tab(1) c_typename(.type) ` ` .name `;`.
end
`} ` camel(.name) `_t;`;
end
You get:
typedef struct {
char *name;
uint16_t age;
} person_t;
Code Generation Templates are a new templating language designed to make it ergonomic to generate source code in C++-like languages. Included with Grain DDL, they are the standard way to maintain code generation backends. Code Generation Templates contains a type system, a standard library of functions, a const expression evaluator and the ability to define your own functions inside of the templates.
However, if Code Generation Templates do not suit your needs, Grain DDL offers a C API, enabling you to walk your data models and types to perform any analysis, presorting or code generation you require. All of this happens at compile time, preventing the need for runtime complexity.
Having a specialized code generation templating system makes code generation maintainable. This is crucial to retaining the investment in your data.
Grain DDL’s types are largely isomorphic to C plain old data types. Grain is designed to produce data that ends up in a typed game programming language like C, C++, Rust, or C#. Philosophically, Grain DDL is much closer to bit and byte manipulation environments than it is a high-level web technology with dictionary-like objects and ill-defined Number types.
Even so, you can use Grain DDL to convert your game data to and from JSON and write serializers for web tech where necessary. It can also be used to specify RESTful interfaces and their models, bypassing a need to rely on OpenAPI tooling.
Crucially, Grain DDL never compromises on the precision required to specify data that runs inside a game engine.
Grain DDL lets you define base data structures, and then derive from them, inheriting defaults where derived fields are not specified. Consider:
blueprint struct BaseArmor {
f32 Fire = 1.0,
f32 Ice = 1.0,
f32 Crush = 1.0,
}
struct WoodArmor copies BaseArmor {
f32 Fire = BaseArmor * 1.5
}
In this case, WoodArmor inherits Ice and Crush at 1.0, but Fire has a +50% increase. This flexible spreadsheet-like approach to building out combat tables for a game gives you a ripple effect on data that immediately permeates all codebases in a project.
Grain DDL is the central definition for your types and data. In some locations, you need data in addition to its name and type to fully specify your type. Consider how health can limited to a range smaller than the type specifies:
struct player {
i8 health = 100
[range min = 0, max = 125]
}
Attributes in Grain DDL are a way of expressing additional data about a field, queryable during code generation. However, it is possible to define an new, strongly typed attribute that represents your bespoke needs:
attr UISlider {
f32 min = -100.0,
f32 max = +100.0,
f32 step = 1.0,
}
struct damage {
f32 amount
[UISlider min = 0.0]
}
This defines a new attribute UISlider which hints that any UI that manipulates damage.amount should use a slider, setting these parameters. Using data inheritance (described above), the slider’s max and step do not change from their defaults, but the minimum is raised to 0.0.
Grain DDL is intended to run as a precompilation step on the developer’s machine, inserting code that is compiled in the target languages of your choice, and with the compiler of your choice. There is no need to link it into shipping executables like a config parsing library. Remove your dependence on runtime reflection.
Grain DDL runs quickly and can generate hundreds of thousands of lines of code in under a second. The software is simply a standalone executable that is under a megabyte. It is is fully supported on Windows, macOS and Linux.
Game developers do not like to slow their compiles down or impose unnecessarily heavy dependencies on developers. Grain DDL is designed to be as lightweight as possible. In practice, Grain replaces multiple code generators, simplifying building.
This article only begins to cover the full featureset of Grain DDL.
As game projects scale, the same data gets defined, processed, and duplicated across a growing number of systems. But game data is more than runtime glue — game data is a long-term asset that outlives game code. It powers ports, benefits from analysis, and extends a title’s shelf life. Grain DDL puts that data under your control, in one place, with one definition — so you can protect and maximize your investment.
Grain DDL is under active development. Email grain@frogtoss.com to get early access, explore example projects, add it to your pipeline.
Fri 09 August 2024 by Michael Labbe
tags code
Sometimes you want to parse a fragment from a string and all you have is C. Parsers for things like rfc3339 timestamps are handy, reusable pieces of code. This post suggests a convention for writing stack-based fragment parsers that can be easily reused or composed into a larger parser.
It’s opinionated, but tends to work for most things so adopt or adapt to your needs.
The idea is pretty simple.
// can be any type
typedef struct {
// fields go here
} type_t;
int parse_type(char **stream, size_t len, type_t *out);
Pass in a **stream pointer to a null-terminated string. On return, **stream points to the location of an error, or past the end of the parse on success. This means that it can point to the null terminator.
Pass in the length of the string to parse to avoid needing to call strlen, or to indicate if the end of a successful parse occurs before the null terminator.
Return can be an int as depicted, or an enum of parse failure reasons if not. The key thing is that zero is success. This allows multiple parses to OR the results and test for error once for trivial code.
That’s the whole interface. You can compose a larger parser out of smaller versions of these. So, if you want to parse a float (a deceptively hard thing to do) in a document, or key value pairs with quotes or something, you can build, test and reuse them by following this convention.
When you implement a fragment parser you end up needing the same few support functions. This suggests a convention.
Testing for whether the stream was fully parsed works well works with a macro containing a single expression:
#define did_fully_parse_stream \
(*stream - start == (ptrdiff_t)len)
int parse_type(char **stream, size_t len, type_t *out) {
char *start = *stream;
if (!did_fully_parse_stream)
return 1;
}
Test the next token for a match:
static int is_token(const char **stream, char ch) {
return **stream == ch;
}
Test the next token and bypass it if it matches. By convention, use this if a token failing to match is not an error.
static int was_token(const char **stream, char ch) {
if (is_token(stream, ch)) {
(*stream)++;
return 1;
}
return 0;
}
Test the next token to be ‘ch’, returning true if it is. While this functionally does the same thing as was_token, it is semantically useful to use it to mean an error has occurred if it does not match.
static int expect_token(const char **stream, char ch) {
return !was_token(stream, ch);
}
Token classification is very easy to implement using C99’s designated initializers. A zero-filled lookup table can be used to test token class and to convert tokens to values.
static char digits[256] = {
['0'] = 0, ['1'] = 1, ['2'] = 2, ['3'] = 3, ['4'] = 4, ['5'] = 5,
['6'] = 6, ['7'] = 7, ['8'] = 8, ['9'] = 9,
};
void func()
{
// is it a digit?
if (digits[**stream]) {
// yes, convert token to stored integral value
int value = digits[**stream];
}
// skip token stream ahead to first non-digit
while (digits[**stream]) (*stream)++;
}
Sun 09 June 2024 by Michael Labbe
tags code
Modern web applications are a façade consisting of many smaller programs and libraries that are configured to run in concert to produce a result. To developers outside of games, this has been acutely obvious for a long time. Games have largely been spared the configuration needs this brings due to a focus on producing a monolithic runtime. However, many modern games ship proprietary logic outside of the code that runs on the disc, such as backend services, so has been affecting games for some time, as well.
At the heart of all this is the need for configuration. Having personally experienced professional devops roles, there seems to be a lack of deep thinking about configuration. This article hopes to inspire deeper thinking about configuration design for programs.
Application configuration is our opportunity to affect runtime state before a program begins its main execution. Static declarations are easily definable, immutable, loggable, can be stored in revision control and can be easily reviewed by a team. Runtime state, on the other hand, is ephemeral and mutable. Through configuration, we have the opportunity wield the runtime state of large, distributed applications in predictable, effcient ways. Most programs do not seize this opportunity.
We treat configuration like it is simple and easy. It is time to start respecting configuration in application design and maintenance.
What is the ground truth configuration for a program? Is it the config file? Not even remotely close. It is the portion of in-memory state that is necessary to cause an (approximately) deterministic, repeatable execution of the program. This is what I call the “ground truth” of an application’s configuration. It usually includes:
Commonly, programs read configuration from many sources. A bespoke search path for configuration, starting from system-wide, and moving in to home directories. Environment variables as an override. Then, command line arguments.
This process differs for each program which is why you’ll see each program document it. Even specifying the system hostname requires addressing multiple files, deprecations and symlinks on Linux.
What happens if there is a system-wide config file but it is not readable because of the permissions of the current user? Pass over it? Throw an error because it exists? This, too, is ambiguous and varies from program to program.
The bottom line is that most programs accumulate a ground truth configuration haphazardly, and then begin executing, perhaps destructively, with no means to review the configuration before it starts.
Writing code is commonly less time consuming than maintaining and debugging the same code. The same is true of configuring software versus troubleshooting it. A misconfigured application produces errors for end users. Many of the configuration formats that are commonly in use (JSON, YAML, TOML) prioritize convenient authorship over unambiguous runtime states. This allows for rapid configuration in exchange for potential risks involving:
Implicit defaults are exceptionally bad when ground truth configuration is not reviewable. You may not even know that you are operating on a bad default, or that an option exists.
Consider:
secrue=true
An insufficiently rigorous program can be misconfigured to breach security without error due to these two aforementioned properties.
YAML, in particular, has a lot of known pitfalls. The point of this article is not to debate popular config file formats. A good developer can overcome YAML’s problems with knowledge and practice, but the problem of contending with underspecified ground truth configuration state is a lifelong drag which can only be overcome through good program design.
JSON, YAML and TOML all have versioned file format specs, but those specifications have no details about how they should affect on-disk performance. Some examples of ambiguities:
Every program behaves differently as a result of this underspecification.
When folks debug a program, they have a mental model of its execution in their heads. Consider:
b = 1;
if (cfg.a)
b += do_optional_thing()
// code continues to do complex things with b
When a developer reads this code, they will either consider b to be augmented by config option a or not. Their mental model of the code necessarily includes this mutating state. Therefore, in removing as much uncertainty as to what the state of a is, is important to someone attempting to ascertain why they are seeing the result of b on their screen.
The rest of this article’s solutions emphasize the need for reducing the size of the mental model necessary for proper configuration troubleshooting.
Which one is right for your application depends on your context. Declarative configuration is a turing-complete program that configures a program. Keeping a mental model of config state requires mentally interpolating variables, simulating loops in your head and jumping through nested function calls.
Imperative configuration lays it all out flat, which lets you see what things are. However, almost everything imperative ends up becoming awkwardly complex when it layers in declarative concepts. See: HCL for_each loops or Ansible adding Python dictionary lookups to YAML files.
A better approach is to think of imperative configuration as a funnel. A data table, perhaps nested, of configuration values can be derived from all sources and fed as input to the ground truth configuration. This table could be declared, or imperatively derived.
The healthy thing is to arrive at a data table of explicit program configuration before core execution of the program starts — an imperative funnel which can be arrived at declaratively.
Schemas are for constraining config file formats, not for constraining ground truth configuration. Ground truth configuration is subjected to underspecified parsers, config file search orders, environment variable and command line overrides and more. Therefore, a schema for a config file does not solve the larger program configuration problem by itself. It doesn’t necessarily hurt it, either, though.
When someone says “we need schemas”, it is useful to explore the root reasoning of that statement before jumping in.
In structured languages, a ground truth configuration can be typed and could be used to produce a schema. The right choice is to keep as much ground truth about the program’s configuration in one reviewable structure.
Most importantly, provide the best tooling for your in-context situation to edit and review the program’s ground truth.
Configuration has a way of becoming layered, especially in devops. For example:
values.yaml file overrides the Helm Chart for a forked Docker imageIn this case, we reap the benefits of a highly-available program that is configured to our specification, compiled by a program provider else and made to work for our purposes. This reduces a large one-time up-front cost. However, we incur a cost of five configuration files, implicitly depending on values from each other to derive whole program state. This has a drag on efficiency for the lifetime of the product. This is an important tradeoff in where you spend effort — one to commit to consciously.
Each small program comes with its own configuration files and state. Since your application consists of multiple programs, you end up producing configuration files that require values that are similar between them. This is brittle when changed.
Further, if there are multiple versions of an application (eg: test and production), there is an n by m problem, where each dependent configuration must exist for each version of the application.
This can be addressed by having a single source of truth for each application configuration, used to produce the smaller configurations for each program.
For the remainder of my life I will depend on large applications that are made up of many small, configured programs operating in concert. Making configuration correct, safe and expressive is an opportunity to wield large numbers of these programs with minimal cost and overhead.
Many of these smaller programs came from programming cultures that emphasized getting something up and running over long term maintainability, loose coupling and quick-and-dirty scripting. As computing complexity increases, it is my hope that the sort of rigorous values that spurred the creation of languages like Rust are applied to configuration management.
Sat 27 May 2023 by Michael Labbe
tags code
Having linear pools of homogenous objects is good for performance, but eventually a piece of code will need to operate on or reserve a single object from the pool. This is a practical reality in gameplay code where things exist for a short time and interact in one-off situations.
What does freeing and allocating a single object from a preallocated pool mean? Let’s clearly establish the scenario:
It is clear to see that allocating and freeing objects from a pool needs to be quick. It happens hundreds if not thousands of times per frame. Linearly searching the list for a free pool entry is not going to yield acceptable performance.
Further, a non-intrusive solution is needed, as the entire array is commonly iterated over during processing to avoid branching and data about freed objects cannot reliably be stored in-array.
The rest of this article explains a constant-time approach to allocating and freeing from the pool.
The Swap Delete technique was first introduced to me in a GDC Canada 2010 talk titled A Dynamic Component Architecture for High Performance Gameplay. It is a talk by Insomniac’s Lead Systems Engineer, Terence Cohen. Outlined in just a few slides in a much larger talk about dynamic component systems, Swap Delete is a brilliant technique that deserves to be plucked out of his larger talk and explained in depth. Note that this technique can be used with any homogenous object pooling system, with or without a component system sitting atop.
First, let’s establish the terminology we need:
Object Pool: An array of homogenous objects.
n is a handle, pool[n] returns the nth object from the object pool. For the purpose of this article, there is no difference between a handle and an index into an array of objects.Roster Table: A table of handles. No handle is repeated, and every handle from the pool is always in the roster table. Handles are unique indexes into the object pool.
Consider this figure: The roster table‘s six values are handles, from 5 to 0. Here, roster[1] returns the handle 4, which is the fourth object in the pool.
Here, partition == 5, which points to the roster table field containing the value 0. Zero is the handle of an unallocated object from the pool.
The partition always points at the next object to allocate. It always separates the handles in the roster table from the allocated ones.
On allocation a handle for the pool object at the partition’s address is returned. The partition is then incremented by 1.
// alloc pseudocode
assert(partition < pool.length)
new_handle = roster_table[partition]
new_obj = pool[new_handle]
partition++
That takes care of constant time allocation, but what about deletion? Clearly, the partition needs to always point at the next free object in the roster table for this allocation scheme to work.
To delete a handle, swap its roster table entry for the entry just before the partition.
In this figure, pool handle 4 from roster index [1] is being deleted. It is swapped into roster index [4], because the partition is pointing at [5].
Now, simply decrement the partition, so that it points at the newly freed roster index entry.
As you can see in the final figure, the partition points at [4], which contains pool handle 4. This is the freshly deleted pool object.
Consider that the next call to alloc will correctly return pool handle 4.
// free pseudocode
roster_index_for_deletion = handle_to_roster_index(handle)
swap(roster_table[roster_index_for_deletion],
roster_table[partition - 1])
partition--
There is a necessary implementation detail that is not mentioned in Terrence Cohen’s presentation and is also not pictured in the graphics in this article.
All calling code will pass objects in by handle, not roster index. The roster table is an implementation detail of the pool.
When an object is freed, it will be passed in by handle, but the roster index is needed to perform the swap. In order to resolve a handle into a roster index, a lookup table from handle to roster index must be created. Its fields must be swapped whenever the roster table’s values are swapped. It is a reverse lookup that must be kept in sync any time the roster table is updated.
That’s it. In total, there are three arrays that must be maintained:
The roster and handle-to-roster tables are implementation details of the pool’s allocation and free functionality; the caller never interacts with them directly. The caller solely deals with handles.
It is possible that a stale handle could point to an object that used to exist but has been freed and re-allocated. This is possible because handles are reused between alloc/free events.
If heap allocated objects were used instead of a pool, stale pointers would likely crash, which is arguably preferable to stale data making its way into a running program which would subtly display strange behaviour.
Andre Weissflog wrote a thought-provoking article on using handles instead of pointers. In an update at the end of his blog post, he proposes adding a generational counter to handles. An object pooling system based on handles would do well to consider generational checks when mapping a handle into a pool, especially while debugging.
Page 1 / 5 »
