The best place to start when building a new agent is its type definitions in its
/sur structure file. The main things to think through are:
- What basic types of data does my agent deal with?
- What actions/commands does my agent need to handle?
- What updates/events will my agent need to send out to subscribers?
- What does my agent need to store in its state?
Let's look at each of these questions in turn, and put together our agent's
/sur file, which we'll call
1. Basic types
Our journal entries will just be plain text, so a simple
@t will work fine to store their contents. Entries will be organized by date, so we'll also need to decide a format for that.
One option would be to use an
@da, and then use the date functions included in the
Date objects. In this case, to keep it simple, we'll just use the number of milliseconds since the Unix Epoch as an
The structure for a journal entry can therefore be:
+$ id @+$ txt @t+$ entry [=id =txt]
Now that we know what a journal entry looks like, we can think about what kind of actions/commands our agent will handle in its
++on-poke arm. For our journal app, there are three basic things we might do:
- Add a new journal entry.
- Edit an existing journal entry.
- Delete an existing journal entry.
We can create a tagged union structure for these actions, like so:
+$ action$% [%add =id =txt][%edit =id =txt][%del =id]==
Updates are a little more complicated than our actions. Firstly, our front-end needs to be able to retrieve an initial list of journal entries to display. Once it has that, it also needs to be notified of any changes. For example, if a new entry is added, it needs to know so it can add it to the list it's displaying. If an entry gets deleted, it needs to remove it from the list. Etc.
The simplest approach to the initial entries is just a
(list entry). Then, for the subsequent updates, we could send out the
$action. Since an
$action is a tagged union, it's simpler to have all updates be a tagged union, so when we get to doing mark conversions we can just switch on the head tag. Therefore, we can define an
$update structure like so:
+$ update$% action[%jrnl list=(list entry)]==
There's one drawback to this structure. Suppose either an agent on a remote ship or an instance of the front-end client is subscribed for updates, and the network connection is disrupted. In the remote ship case, Gall will only allow so many undelivered messages to accumulate in Ames before it automatically kicks the unresponsive subscriber. In the front-end case, the subscription will also be ended if enough unacknowledged messages accumulate, and additionally the client may sometimes need to establish an entirely new connection with the ship, discarding existing subscriptions. When this happens, the remote ship or web client has no way to know how many (if any) updates they've missed.
The only way to resynchronize their state with ours is to discard their existing state, refetch the entire initial state once again, and then resubscribe for updates. This might be fine if the state of our agent is small, but it becomes a problem if it's very large. For example, if our agent holds tens of thousands of chat messages, having to resend them all every time anyone has connectivity issues is quite inefficient.
One solution to this is to keep an update log. Each update can be tagged with the time it occurred, and stored in our agent's state, separately to the entries. If an agent or web client needs to resynchronize with our agent, it can just request all updates since the last one it received. This approach is used by the
%graph-store agent, for example. Our agent is local-only and doesn't have a huge state so it might not be strictly necessary, but we'll use it to demonstrate the approach.
We can define a logged update like so, where the
@ is the update timestamp in milliseconds since the Unix Epoch:
+$ logged (pair @ action)+$ update%+ pair @$% action[%jrnl list=(list entry)][%logs list=(list logged)]==
We need to store two things in our state: the journal entries and the update log. We could just use a couple of
maps like so:
+$ journal (map id txt)+$ log (map @ action)
maps are fine if we just want to access one value at a time, but we want to be able to:
- Retrieve only some of the journal entries at a time, so we can have "lazy loading" in the front-end, loading more entries each time the user scrolls to the bottom of the list.
- Retrieve only logged updates newer than a certain time, in the case where the subscription is interrupted due to connectivity issues.
- Retrieve journal entries between two dates.
Maps are ordered by the hash of their key, so if we convert them to a list they'll come out in seemingly random order. That means we'd have to convert the map to a list, sort the list, and then iterate over it again to pull out the items we want. We could alternatively store things in a list directly, but retrieving or modifying arbitrary items would be less efficient.
To solve this, rather than using a
map or a
list, we can use an ordered map. The mold builder for an ordered map is a
mop, and it's included in the
zuse.hoon utility library rather than the standard library.
mop is defined similarly to a
map, but it takes an extra argument in the following manner:
((mop key-mold val-mold) comparator-gate)
The gate is a binary gate which takes two keys and produces a
?. The comparator is used to decide how to order the items in the mop. In our case, we'll create a
mop like so:
+$ journal ((mop id txt) gth)+$ log ((mop @ action) lth)
The entries in
$journal are arranged in ascending time order using
++gth, so the right-most item is the newest. The
mop contains the update log, and is arranged in descending time order, so the right-most item is the oldest.
We'll look at how to use ordered maps later when we get to writing the agent itself.
When we put each of these parts together, we have our complete
Click to expand
|%:: Basic types of the data we're dealing with::+$ id @+$ txt @t+$ entry [=id =txt]:: Poke actions::+$ action$% [%add =id =txt][%edit =id =txt][%del =id]==:: Types for updates to subscribers or returned via scries::+$ logged (pair @ action)+$ update%+ pair @$% action[%jrnl list=(list entry)][%logs list=(list logged)]==:: Types for our agent's state::+$ journal ((mop id txt) gth)+$ log ((mop @ action) lth)--
App School I /sur section - This section of App School covers writing a
/surstructure library for an agent.
Ordered map functions in
zuse.hoon- This section of
zuse.hooncontains all the functions for working with
mops, and is well commented.