Skip to Content

TLA+ Action Properties

There’s not a whole lot on TLA+ technique out there: all the resources are either introductions or case studies. Good for people starting out, bad for people past that. I think we need to write more intermediate-level stuff, what Ben Kuhn calls Blub studies. Here’s an attempt at that.

Most TLA+ properties are invariants, properties that must be true for every state in the behavior. If we have a simple counter:

EXTENDS Integers
VARIABLES x

Init == 
  x = 1

Next == 
  /\ x < 3
  /\ x' = x + 1

Spec == Init /\ [][Next]_x

Then an invariant of this spec is x \in 1..3. One property that is not an invariant is that x is always increasing. Going from x = 2 to x = 1 would violate this, but those are both individually valid states. It’s only the transition that is invalid. This property on a transition is called an action property and is written [][x' > x]_x.

show syntax

(x' > x) is the statement that the value of x in the next state (x') is greater than the value of x in the current state. [](x' > x) would say that this is always true: every state must have a greater value for x than the state before.

We cannot write that directly, since it would be violated by stutter steps. Since TLA+ is stutter-invariant, we should be able to insert a stutter anywhere without breaking the property. We instead want [](x' > x \/ UNCHANGED x), that either x is increasing or doesn’t change. TLA+ provides the shorthand [][A]_x syntax, finally giving us [][x' > x]_x.

In the Toolbox, action properties go in the “Temporal Properties” box. If you’re running from the command line, they’re PROPERTYs in the config file.

Use Cases

Conditional Properties

Using =>, we can make “conditional” properties that only must hold if the precondition is true. For example, if our system has a “kill switch”, we can say some value should not change while system is disabled:1

[][disabled => UNCHANGED x]_<<disabled, x>>

Or we can say that certain things must change in lockstep:

\* If x changes, y must become the old value of x
[][x' /= x => y' = x]_<<x, y>>

We can also use actions as the preconditions for invariants, where we only need the invariant to hold under those specific actions. If we split our spec into Machine and World actions, we might want that only Machine actions maintain the invariant:

\* Next == Machine \/ World
\* Inv is some invariant
[][Machine => Inv]_vars

We can use actions on both sides of the condition, such as to say that if a certain change happens, it must have been “because of” a certain action:

\* owner \in [Credit -> User]
\* offers \in SUBSET (User \X User \X Credit)

Accept(from, to, credit) ==
\*    (User, User, Credit)
  /\ <<from, to, credit>> \in offers
  /\ offers' = offers \ {<<from, to, credit>>}
  /\ owner' = [owner EXCEPT ![credit] = to]

\* If ownership changes from A to B
\* It's because B accepted an offer from A
ValidChange(credit) ==
  LET co == owner[credit]
  IN co /= co' =>
    Accept(co, co', credit)

\* All changes in the system are valid changes   
ChangeProp ==
  [][\A c \in Credits: ValidChange(c)]_owner 

State Transitions

A server has three states: Offline, Booting, and Online. We can say the server cannot go directly from Offline to Online:

\* status \in {Offline, Booting, Online}
[][status = Offline => status' /= Online]_status

If we have many different possible state transitions, we can abstract the valid transitions into an operator:

\* States == {A, B, C, D}
\* state \in States
Transitions == {
  <<A, B>>, <<A, C>>
  <<B, D>>,
  <<C, A>>, <<C, B>>
}

[][<<state, state'>> \in Transitions]_state

We can combine transitions with conditional properties:

T == <<state, state'>>
[][T = <<A, B>> => x' > x]_<<state, x>>
[][x' < x => T = <<A, C>>]_<<state, x>>

And we can extend state transition properties to multiple concurrent state machines:

\* Machines == {M1, M2, M3}
\* state \in [Machines -> States]

ChangedState(m) == state[m] /= state'[m]
TransitionAction(m) == <<state[m], state'[m]>>
[][\A m \in Machines:
  ChangedState(m) => 
    TransitionAction(m) \in Transitions
]_state

Other Uses

  • We are sharing a lock between threads. If one thread holds the lock, it must release the lock before another thread can acquire it:2

    \* lock \in Threads \union {NULL}
    [][lock /= NULL => lock' \in {lock, NULL}]_lock
    
  • We have an event log that is append-only. Prior values of the log should never change. This means that log is a prefix of log':

    [][SubSeq(log', 1, Len(log)) = log]_log 
    
  • Once a predicate becomes true, it cannot become false again:

    \* flag \in BOOLEAN
    [][flag => flag']_flag
    

    Once a value is no longer NULL, it does not go back to NULL:

    [][val /= NULL => val' /= NULL]_val
    

    Once a value is no longer NULL, it doesn’t change:

    [][val /= NULL => UNCHANGED val]_val
    
  • If we’re processing messages from a client and there are still messages in their queue, we do not switch to another client:

    \* curr = current_client \in Clients
    \* queue \in [Clients -> Seq(Msg)]
    [][
      queue[curr] /= <<>> => UNCHANGED curr
    ]_<<curr, queue>>
    

Notes

  • In my experience, action properties are less common than invariants and more common than liveness properties. However, liveness properties are generally “more important” to a spec being valid than the action properties are.3 Action properties are very useful, but not critical in the same way.

  • The x in [A]_x can be any state predicate, not just a variable. Usually people add a helper operator vars == <<x, y, z, ...>> and then write [A]_vars. The only requirement is that all variables appearing in A must also appear in the subscript.

  • Action properties can only describe pairs of states. You can’t natively write a property that spans three states without adding an auxilary variable.

  • There’s a second form of action syntax: <<A>>_x means A /\ x' /= x, or “A happens and x changes.” Just as we can check “eventually predicate P is true” with <>P, we can express “eventually action A happens” with <><<A>>_x. Unfortunately, TLC cannot check this. It can only check properties of the form []<><<A>>_x, or “A happens infinitely often”.

  • TLC can also check properties of the form <>[][A]_x, or “eventually, the action property [][A]_x always holds.” I used that to form a conditional property in this post:

    Resilient == []<>Safe
    RareWorldResilient == <>[][World => Safe]_x => Resilient
    

    That property, in English, is “If it’s eventually the case that World actions only happen when Safe holds, then even if World breaks Safe, we will always eventually return to Safe holding.”

    I don’t think I’ve actually used <>[][A]_x in a real world spec, though, just in examples.

  • Most specs are written as Init /\ [][Next]_vars. TLC can check [][Next]_vars just like any other action property! This is the launching point into refinement, or using an entire specification as the checked property of another spec. That’ll have to be the topic of another post.

Thanks to Andrew Helwer for feedback.


  1. UNCHANGED x is syntactic sugar for x' = x. [return]
  2. I could have just written lock' = NULL instead of lock' \in {lock, NULL}. [P]_lock means P \/ UNCHANGED lock, which would have implicitly permitted the thread lock to not change. But that would “offload” spec logic to the syntax, which isn’t clear to readers. [return]
  3. Invariants and action properties only cover bad states and transitions. They don’t cover “the system actually does what you need it to do.” That’s liveness. [return]