This document describes Gall, the Arvo userspace vane.
Agents
Urbit code lives in the following basic categories:
- Runtime (Nock interpreter, persistence engine, IO drivers, jets)
- Kernel vanes (managed by Arvo)
- Userspace agents (managed by Gall, permanent state)
- Userspace threads (managed by Spider, transient state)
This lays out the framework for the third category: Agents.
An agent is a piece of software that is primarily focused on maintaining and distributing a piece of state with a defined structure. It exposes an interface that lets programs read, subscribe to, and manipulate the state. Every event happens in an atomic transaction, so the state is never inconsistent. Since the state is permanent, when the agent is upgraded with a change to the structure of the state, the developer provides a migration function from the old state type to the new state type.
It's not too far off to think of an agent as simply a database with developer-defined logic. But an agent is significantly less constrained than a database. Databases are usually tightly constrained in one or more ways because they need to provide certain guarantees (like atomicity) or optimizations (like indexes). Urbit is a single-level store, so atomicity comes for free. Many applications don't use databases because they need relational indices; rather, they use them for their guarantees around persistence. Some do need the indices, though, and it's not hard to imagine an agent which provides a SQL-like interface.
On the other hand, an agent is also a lot like what many systems call a "service". An agent is permanent and addressable -- a running program can talk to an agent just by naming it. An agent can perform IO, unlike most databases. This is a critical part of an agent: it performs IO along the same transaction boundaries as changes to its state, so if an effect happens, you know that the associated state change has happened. You should be careful, though, to avoid creating implicit non-atomicity -- eg if it comes into an inconsistent state pending a response to an IO action.
But the best way to think about an agent is as a state machine. Like a state machine, any input could happen at any time, and it must react coherently to that input. Ouput (effects) and the next state of the machine are both pure functions of the previous state and the input event. Of course, it's important to ensure there isn't an order of events that could cause your agent to enter an inconsistent state.
We often think of state machines as finite state machines, but of course in this instance, the state is not actually finite (though it should be definable in a regular recursive data type).
Specification
An agent is defined as a core with a set of arms to handle various events. These handlers usually produce a list of effects and the next state of the agent. The interface definition can be found in sys/lull.hoon, which at the time of writing is:
++ agent=< form|%+$ step (quip card form)+$ card (wind note gift)+$ note$% [%agent [=ship name=term] =task][%arvo note-arvo][%pyre =tang]::[%grow =spur =page][%tomb =case =spur][%cull =case =spur]==+$ task$% [%watch =path][%watch-as =mark =path][%leave ~][%poke =cage][%poke-as =mark =cage]==+$ gift$% [%fact paths=(list path) =cage][%kick paths=(list path) ship=(unit ship)][%watch-ack p=(unit tang)][%poke-ack p=(unit tang)]==+$ sign$% [%poke-ack p=(unit tang)][%watch-ack p=(unit tang)][%fact =cage][%kick ~]==++ form$_ ^||_ bowl++ on-init*(quip card _^|(..on-init))::++ on-save*vase::++ on-load|~ old-state=vase*(quip card _^|(..on-init))::++ on-poke|~ [mark vase]*(quip card _^|(..on-init))::++ on-watch|~ path*(quip card _^|(..on-init))::++ on-leave|~ path*(quip card _^|(..on-init))::++ on-peek|~ path*(unit (unit cage))::++ on-agent|~ [wire sign]*(quip card _^|(..on-init))::++ on-arvo|~ [wire sign-arvo]*(quip card _^|(..on-init))::++ on-fail|~ [term tang]*(quip card _^|(..on-init))----
Here's a skeleton example of an implementation:
^- agent:gall=| state=@|_ bowl:gall+* this .++ on-init`this::++ on-save!>(state)::++ on-load|= =old-state=vase`this(state !<(@ old-state-vase))::++ on-poke|= [=mark =vase]~& > state=state~& got-poked-with-data=mark=. state +(state)`this::++ on-watch|= path`this::++ on-leave|= path`this::++ on-peek|= path*(unit (unit cage))::++ on-agent|= [wire sign:agent:gall]`this::++ on-arvo|= [wire sign-arvo]`this::++ on-fail|= [term tang]`this--
We also supply a default-agent library, which is useful if you don't want to handle some of the arms. Most of the above could also be implemented as:
/+ default-agent^- agent:gall=| state=@|_ =bowl:gall+* this .default ~(. (default-agent this %|) bowl)::++ on-init on-init:default++ on-save on-save:default++ on-load on-load:default++ on-poke|= [=mark =vase]~& > state=state~& got-poked-with-data=mark=. state +(state)`this::++ on-watch on-watch:default++ on-leave on-leave:default++ on-peek on-peek:default++ on-agent on-agent:default++ on-arvo on-arvo:default++ on-fail on-fail:default--
So, an agent is a core with 10 arms. The handlers correspond to different sorts of input. We'll discuss each of these in detail, but first a few general concepts.
Basic concepts
Interacting with an agent
There are two basic ways programs can interact with an agent: you may subscribe to data as described above, or you may "poke" an agent to send it a command. Suppose there is an agent for a calendar service; then, pokes would likely include %create-event, %modify-event, %delete-event, and %change-time-zone (along with assocated data). Subscription paths could include /next-event and /all-events.
Agents generally conform to CQRS -- pokes may change state, but subscriptions generally shouldn't. In practice, there may be state changes required to properly initialize the subscription, but these shouldn't change the essential state of the agent.
Cards
Most of the handlers produce a (quip card agent:gall), which is just [(list card) agent:gall]. The first allows us to produce effects, and the second allows us to maintain state.
A card is one of two things: a note or a gift. You %pass notes and you %give gifts. When you pass a note, you expect zero or more responses; when you give a gift, you will not get a response. Since you may get a response to a note, you tag it with a wire so that you can identify the response.
Thus, we say you "pass a note along a wire", or you "give a gift". These phrases correspond to producing a card that looks like:
[%pass /my/wire a-note]
or
[%give a-gift]
When you give a gift, that's the end of the story, but when you pass a note, you may get a response. If you do, then the response will come tagged with the wire that you used to pass the note. It's generally good practice to make your wires fairly unique within your agent, since otherwise you may not be able to distinguish the responses.
Notes
An agent may pass notes to either Arvo or another agent. If the note is to another agent, then it should usually be one of these:
[%pass /my/wire %agent [our.bowl agent-name] %watch /a/path][%pass /my/wire %agent [our.bowl agent-name] %leave ~][%pass /my/wire %agent [our.bowl agent-name] %poke %foo-mark !>(poke-data)]
Note that to unsubscribe to a path, you must send the unsubscription on the same wire that sent for the original subscription. Don't subscribe to separate paths along the same wire, because then you can't properly distinguish them for cancelling (besides not being able to distinguish the subscription updates). In other words, besides letting you distinguish your cards, the wire also identifies requests for the system.
If the note is not to another agent but to Arvo itself, then it is a request to one of the vanes, which are kernel modules that provide various system services, including IO. Here are some examples:
[%pass /my/wire %arvo %b %wait (add now.hid ~s10)]
This is a request to the %b vane (Behn, the timer vane) to set a timer for 10 seconds in the future. After 10 seconds, Behn will respond with a %wake card, which you will recieve in +on-arvo. It will come back on /my/wire.
=/ =path /(scot %p our.hid)/base/(scot %da now.hid)/my-file/txt=/ contents=cage [%txt !>(~['text file line 1' 'line 2'])][%pass /my/wire %arvo %c %info (foal:space:userlib path contents)]
This is a request to the %c vane (Clay, the filesystem vane) to write a file to /my-file/txt in the base desk. You will not receive a response to this note.
=/ =path /my-file/txt=/ =moat [da+now.hid da+(add now.hid ~h1) path][%pass /my/wire %arvo %c %warp our.hid %base ~ %many & moat]
This is a request to the %c vane to send us a %writ card whenever /my-file/txt changes in the next hour. There may be many responses to this note if the file changes many times.
=/ =request:http[%'GET' 'https://example.com' ~ ~][%pass /my/wire %arvo %i %request request *outbound-config:iris]
This is a request to the %i vane (Iris, the HTTP client vane) to make a GET HTTP request to example.com↗. The response will come as an %http-response card.
Gifts
In contrast to the many possible notes, there are only two types of gifts that an agent may give:
[%give %fact (unit path) =cage][%give %kick (unit path)]
A subscription update is a new piece of subscription content for all subscribers on a given path. If no path is given, then the update is only given to the program that instigated this request. Typical use of this mode is in +on-watch to give an initial update to a new subscriber to get them up to date.
A subscription close closes the subscription for all subscribers on a given path. If no path is given, then the update is only given to the program that instigated this request. Typical use of this mode would be in +on-watch to produce a single update to a subscription then close the subscription.
Vases and cages
A vase is a piece of dynamic data. Structurally, it's a pair of an explicit reification of a type and an untyped noun. This lets us represent a value which has a type that isn't known at compile time. A vase has three operations:
!>is a unary rune that lifts a statically typed value to a dynamically-typedvase. For example,!>('hi')gives[#t/@t q=26.984].!<is a binary rune that takes a mold and a dynamically typedvaseand reduces it to a statically typed value. If thevasedoes not in fact have the type of the mold you give it, then it nest-fails, else it producesvalue. For example,!<(@t !>('hi'))produces'hi'while!<(^ !>('hi'))causes a nest fail..The compiler takes text and converts it to a
vaseof the compiled code. Agents shouldn't need this directly, but Gall uses this to compile agents tovases, on which it calls!<(agent:gall compiled-agent-vase).
A cage is simply a pair of a mark and a vase. A mark is a textual tag that should correspond to the particular dynamic type in the vase.
In agents, we use vases to represent types which Gall doesn't know about when it was compiled, but which nevertheless need to go outside the agent. In practice, there are three common cases:
The data in a "poke" request is of a type that is defined by each agent, so it must by dynamic. This is the input to
++on-pokeas well as the cage in the%pokecase of the%agentnote.The data in a subscription update is defined by each agent, so it must be dynamic. This is the cage in
%fact, both when giving the update and in the input to+on-agent.The state of an agent is also unique to each agent. Most of the time, Gall doesn't interact directly with agent's state, but when upgrading an agent, it must pass the state of the old version of the agent to the new version of the agent. This is the output of
+on-saveand the input to+on-load.
State
In the definition of an agent, the entire core is wrapped with the ^| rune, which corresponds to the "iron" variance mode. This means it's "contravariant" in the input, which allows Gall to write to the new bowl on every event. It also means that the context of the core is opaque to Gall, which means you can store state in your context in whatever structure you want.
This is why most of the handler arms produce not just a list of cards but also a new agent:gall. Usually, this will be the "same" agent in the sense that the code will be the same, but you may have changed the state that's stored in its (opaque to Gall) context.
The skeleton example above gives an example of keeping an atom as your state. Define the state in the context of your agent core, then you can change it with =. or %=, as long as you produce the new version of the agent core.
When you upgrade an agent, you need to extract the state from your opaque context and produce it to Gall as a dynamically typed vase. Usually, this will be easy: just call !> on whatever state you wish to preserve. This is +on-save.
When the new agent is about to be started, Gall will call it with +on-load with the vase just produced above. This allows you to ingest your old state and continue right where you left off.
If the type of your state hasn't changed, you can just
=. state !<(state-type old-state-vase)
If it has changed, then it should look more like:
=. state (upgrade-state !<(old-state-type old-state-vase))
It's useful to tag your state with a version number so that the +upgrade-state function can take a tagged union of all your old state types and upgrade from any of them to the current state type. For example, it may have the following structure:
+$ old-state-types$% [%0 s=state-0][%1 s=state-1][%2 s=state-2]==++ upgrade-state|= old=old-state-types=? old ?=(%0 -.old)(state-0-to-1 s.old)=? old ?=(%1 -.old)(state-1-to-2 s.old)?> ?=(%2 -.old)s.old
Bowl
The core takes as input a bowl, which includes useful info like the current ship, the current time, and a renewable source of entropy. This information is available to any of the handlers.
Arms
A description of each of the handler arms follows.
+on-init
This arm is called once when the agent is started. It has no input and lets you perform any initial IO.
+on-save
This arm is called immediately before the agent is upgraded. It packages the permament state of the agent in a vase for the next version of the agent. Unlike most handlers, this cannot produce effects.
+on-load
This arm is called immediately after the agent is upgraded. It receives a vase of the state of the previously-running version of the agent, which allows it to cleanly upgrade from the old agent.
+on-poke
This arm is called when the agent is "poked". The input is a cage, so it's a pair of a mark and a dynamic vase.
+on-watch
This arm is called when a program wants to subscribe to the agent on a particular path. The agent may or may not need to perform setup steps to intialize the subscription. It may produce a %give %subscription-result to the subscriber to get it up to date, but after this event is complete, it cannot give further updates to a specific subscriber. It must give all further updates to all subscribers on a specific path.
If this arm crashes, then the subscription is immediately terminated. More specifcally, it never started -- the subscriber will receive a negative %watch-ack. You may also produce an explicit %kick to close the subscription without crashing -- for example, you could produce a single update followed by a %kick.
+on-leave
This arm is called when a program becomes unsubscribed to you. Subscriptions may close because the subscriber intentionally unsubscribed, but they also could be closed by an intermediary. For example, if a subscription is from another ship which is currently unreachable, Ames may choose to close the subscription to avoid queueing updates indefinitely. If the program crashes while processing an update, this may also generate an unsubscription. You should consider subscriptions to be closable at any time.
+on-peek
This arm is called when a program reads from the agent's "scry" namespace, which should be referentially transparent. Unlike most handlers, this cannot perform IO, and it cannot change the state. All it can do is produce a piece of data to the caller, or not.
If this arm produces [~ ~ data], then data is the value at the the given path. If it produces [~ ~], then there is no data at the given path and never will be. If it produces ~, then we don't know yet whether there is or will be data at the given path.
+on-agent
This arm is called to handle responses to %give moves to other agents. It will be one of the following types of response:
%poke-ack: acknowledgment (positive or negative) of a poke. If the value is~, then the poke succeeded. If the value is[~ tang], then the poke failed, and a printable explanation (e.g. a stack trace) is given in thetang.%watch-ack: acknowledgment (positive or negative) of a subscription. If negative, the subscription is already ended (technically, it never started).%fact: update from the publisher.%kick: notification that the subscription has ended.
+on-arvo
This arm is called to handle responses to %pass moves to vanes. The list of possible responses from the system is statically defined in sys/lull.hoon (grep for +$ sign-arvo).
+on-fail
If an error happens in +on-poke, the crash report goes into the %poke-ack response. Similarly, if an error happens in +on-subscription, the crash report goes into the %watch-ack response. If a crash happens in any of the other handlers, the report is passed into this arm.