Some questions about inter-canister call

[baolongt]

Hi Devs! I have some question about inter-canister call

1. How many consensus rounds does a typical inter-canister call require? (caller → target →
   caller response)
2. What’s the latency difference when:
   • Both canisters on same subnet?
   • Canisters on different subnets?
3. If target canister exposes a query method (like icrc1_balance_of), does calling
   it from an update context still require full consensus?
4. What’s the theoretical minimum latency for:
   • Same-subnet inter-canister call?
   • Cross-subnet inter-canister call?

[C-B-Elite]

@baolongt

How many consensus rounds does a typical inter-canister call require? (caller → target →
caller response)

If the canisters are on the same subnet, a canister-to-canister call may complete within the same execution round (each consensus round includes one execution round).
If they are not on the same subnet, based on my understanding, it requires at least four consensus rounds.

@baolongt

What’s the latency difference when:

• Both canisters on same subnet?
• Canisters on different subnets?

Inter-canister call latency on the same subnet is significantly lower than any inter-canister call across different subnets. Calls between canisters on different subnets are referred to as X-Net calls, which operate at the level of communication between two subnets and are constrained by factors such as subnet signatures and subnet queues.

@baolongt

If target canister exposes a query method (like icrc1_balance_of), does calling
it from an update context still require full consensus?

Canister-to-canister calls are always update calls, regardless of whether the target method is declared as a query or an update. In practice, latency is influenced much more by whether the canisters are on the same subnet than by whether the call is an update or a query.

@baolongt

What’s the theoretical minimum latency for:

• Same-subnet inter-canister call?
• Cross-subnet inter-canister call?

You can find more information about latency here:
https://learn.internetcomputer.org/hc/en-us/articles/39320190051348-Performance

[Vivienne]

@C-B-Elite

If the canisters are on the same subnet, a canister-to-canister call may complete within the same execution round (each consensus round includes one execution round).

This is correct, although this will only work if source and target canister have very little load on them. With low load I would expect 1 round to call the target canister and 1 round to start processing the response. With high load it will take longer

@C-B-Elite

If they are not on the same subnet, based on my understanding, it requires at least three consensus rounds.

Four rounds actually:

• 1 round to put the call in the output queue of the source subnet
• 1 round for the target subnet to accept the call
• 1 round for the target subnet to run the call and put it in its output queue
• 1 round for the source subnet to accept the call

@C-B-Elite

@baolongt

If target canister exposes a query method (like icrc1_balance_of), does calling
it from an update context still require full consensus?

Canister-to-canister calls are always update calls, regardless of whether the target method is declared as a query or an update. In practice, latency is influenced much more by whether the canisters are on the same subnet than by whether the call is an update or a query.

Correct in this context, small addition: canister-to-canister calls can also be of type ‘composite query’, which doesn’t run under consensus. But then the call is not made from an update context