Quantify and expand coverage of TLC simulation mode

[lemmy]

TLC's simulation mode is generally effective at detecting regressions while being less resource-intensive than full model checking, which helps reduce the strain on our CI resources. However, the actual coverage of simulation mode is unknown. Additionally, increasing coverage is highly desirable.

• Quantify current coverage
• Check liveness properties during simulation
  • Check high-level liveness properties in simulation mode #6545
• Expand coverage by starting simulation from a larger set of initial states
  • Manually defined set of initial states
  • Adjust probabilities of failure actions based on length of trace, ...
  • RandomSetOfSubsets inductive invariant candidate

[lemmy]

Quantify current coverage with five nodes

Main observation: The RcvProposeVoteRequest action rarely, if ever, occurs, likely because the upper limit set for the length of the generated traces (-depth) is too restrictive. Note that exhaustive model checking shows that it takes 28/47 states with 2/3 nodes and 2 configuration for one of th nodes to be Retired Committed (client requests disabled).

The following plots display statistics gathered from a long-running simulation (note that the overall overhead associated with extended statistics (-Dtlc2.tool.Simulator.extendedStatistics=true) and the accuracy of the probabilistic data structure have been analyzed independently).

Python code used to generate the plot

import matplotlib.pyplot as plt
import pandas as pd
data = pd.read_json("SIMccfraft.ndjson", lines=True)
json_data = data.iloc[-1].to_dict()
actions_df = pd.DataFrame(json_data["actions"].items(), columns=["Action", "Frequency"])
actions_df_sorted = actions_df.sort_values(by="Frequency", ascending=False)
plt.figure(figsize=(10, 6))
plt.bar(actions_df_sorted["Action"], actions_df_sorted["Frequency"])
plt.yscale('log')
plt.xticks(rotation=90)
plt.xlabel('Actions')
plt.ylabel('Frequency (Log scale)')
plt.title('Action Frequencies with Logarithmic Y-Axis')
plt.tight_layout()
plt.show()

Python code used to generate the plot

import matplotlib.pyplot as plt
import pandas as pd
data = pd.read_json("SIMccfraft.ndjson", lines=True)
plt.plot(data['duration'], data['distinct'], marker='x', linestyle='--', color='black', label='distinct')
plt.plot(data['duration'], data['generated'], marker='o', linestyle='-', color='blue', label='generated')
plt.title('Distinct states over Duration')
plt.xlabel('Duration')
plt.ylabel('Count')
plt.grid(True)
plt.legend()
plt.show()

Variables with finitely many distinct values:

• membershipState Cardinality([Servers -> MembershipState])
• leadershipState Cardinality([Servers -> LeeadershipState])
• retirementCompleted Cardinality([Servers -> SUBSET Servers])
• hasJoined Cardinality([Servers -> BOOLEAN])
• isNewFollower Cardinality([Servers -> BOOLEAN])

Python code used to generate the plot

import matplotlib.pyplot as plt
import pandas as pd
import matplotlib.cm as cm
import numpy as np
data = pd.read_json("SIMccfraft.ndjson", lines=True)
plt.figure(figsize=(10, 6))
i = 1
colors = cm.rainbow(np.linspace(0, 1, len(data['distinctvalues'].iloc[-1].keys())))
for value, color in zip(data['distinctvalues'].iloc[-1].keys(),colors):
    plt.subplot(4, 4, i)
    plt.plot(data['duration'], data['distinctvalues'].apply(lambda x: x.get(value, None)), marker='o', linestyle='-', label=value, color=color)
    plt.xlabel('Duration')
    plt.ylabel('Count')
    plt.grid(True)
    plt.legend()
    i += 1
plt.show()

@lemmy ➜ /workspaces/CCF (main) $ git diff
diff --git a/tla/consensus/SIMccfraft.cfg b/tla/consensus/SIMccfraft.cfg
index 675b9e322..de6948a74 100644
--- a/tla/consensus/SIMccfraft.cfg
+++ b/tla/consensus/SIMccfraft.cfg
@@ -49,12 +49,15 @@ CONSTANTS
     Extend <- [abs]ABSExtend
     CopyMaxAndExtend <- [abs]ABSCopyMaxAndExtend
 
-CONSTRAINT
-    StopAfter
+\* CONSTRAINT
+\*     StopAfter
 
 CHECK_DEADLOCK
     FALSE
 
+_PERIODIC
+    PrintStats
+
 PROPERTIES 
     CommittedLogAppendOnlyProp
     MonotonicTermProp
diff --git a/tla/consensus/SIMccfraft.tla b/tla/consensus/SIMccfraft.tla
index 3d86eb251..5663f08d2 100644
--- a/tla/consensus/SIMccfraft.tla
+++ b/tla/consensus/SIMccfraft.tla
@@ -45,6 +45,9 @@ DebugInvUpToDepth ==
     \* -depth after generating the first trace.
     TLCGet("level") < TLCGet("config").depth
 
+PrintStats ==
+    Serialize(<<TLCGet("stats")>>, "SIMccfraft.ndjson", [format |-> "NDJSON", charset |-> "UTF-8", openOptions |-> <<"WRITE", "CREATE", "APPEND">>])
+
 ----
 \* Refinement

---

Slightly better coverage with -depth 1000:

[achamayou]

@lemmy a detail, but:

plt.plot(data['duration'], data['generated'], marker='o', linestyle='-', color='blue', label='distinct')

should be:

plt.plot(data['duration'], data['generated'], marker='o', linestyle='-', color='blue', label='generated')

[achamayou]

@lemmy increasing the depth does sound like a good step to improve coverage.

RequestVote seems overly represented, even if we think that we want to focus on fault-heavy scenarios. Similarly, it looks like (and this is very rough from eyeballing the graph) that request votes is about two orders of magnitude more frequent than BecomeLeader. That's a lot of failed elections!

My sense is that we should aim to diminish timeout frequency, by maybe an order of magnitude or so, to get to stage where we are generally progressing further. It's still be quite a lot more faults than a typical system.

[lemmy]

I've been running simulations of the specification for a few hours, and so far, neither LogMatchingProp nor LeaderProp have been violated. However, the overall coverage is limited (see below) because I had to remove RandomElement, as it doesn't work well ENABLED/fairness. The incompatibility of RandomElement and ENABLED suggests to revive reinforcement learning to a) eliminate RandomElement from the spec, and b) learn the function that increases coverage automatically.

@lemmy ➜ /workspaces/CCF (main) $ git diff
diff --git a/tla/consensus/SIMccfraft.cfg b/tla/consensus/SIMccfraft.cfg
index d09154097..db75ab5b0 100644
--- a/tla/consensus/SIMccfraft.cfg
+++ b/tla/consensus/SIMccfraft.cfg
@@ -48,8 +48,8 @@ CONSTANTS
     Extend <- [abs]ABSExtend
     CopyMaxAndExtend <- [abs]ABSCopyMaxAndExtend
 
-CONSTRAINT
-    StopAfter
+\* CONSTRAINT
+\*     StopAfter
 
 CHECK_DEADLOCK
     FALSE
@@ -64,6 +64,9 @@ PROPERTIES
     PendingBecomesFollowerProp
     NeverCommitEntryPrevTermsProp
     RefinementToAbsProp
+    
+    LogMatchingProp
+    LeaderProp
 
 \* ALIAS
 \*     \* DebugAlias
diff --git a/tla/consensus/SIMccfraft.tla b/tla/consensus/SIMccfraft.tla
index 0c504749b..a4a3cbdc4 100644
--- a/tla/consensus/SIMccfraft.tla
+++ b/tla/consensus/SIMccfraft.tla
@@ -17,20 +17,12 @@ SIMInitReconfigurationVars ==
     \/ CCF!InitReconfigurationVars
 
 SIMCheckQuorum(i) ==
-    /\ 1 = RandomElement(1..10)
     /\ CCF!CheckQuorum(i)
 
 SIMChangeConfigurationInt(i, newConfiguration) ==
-    /\ 1 = RandomElement(1..100)
     /\ CCF!ChangeConfigurationInt(i, newConfiguration)
 
 SIMTimeout(i) ==
-    /\ \/ 1 = RandomElement(1..100)
-       \* Always allow Timeout if no messages are in the network
-       \* and no node is a candidate or leader.  Otherwise, the system
-       \* will deadlock if 1 # RandomElement(...).
-       \/ /\ \A s \in Servers: leadershipState[s] \notin {Leader, Candidate}
-          /\ Network!Messages = {}
     /\ CCF!Timeout(i)
 
 \* The state constraint  StopAfter  stops TLC after the alloted
diff --git a/tla/consensus/ccfraft.tla b/tla/consensus/ccfraft.tla
index 4857ff4b9..16d6e5d9d 100644
--- a/tla/consensus/ccfraft.tla
+++ b/tla/consensus/ccfraft.tla
@@ -1298,6 +1298,8 @@ Spec ==
     /\ \A s \in Servers : WF_vars(AdvanceCommitIndex(s))
     /\ \A s \in Servers : WF_vars(BecomeLeader(s))
     /\ \A s \in Servers : WF_vars(Timeout(s))
+    /\ \A s \in Servers : WF_vars(AppendRetiredCommitted(s))
+    /\ \A s \in Servers : WF_vars(ChangeConfiguration(s))
 
 ------------------------------------------------------------------------------
 \* Correctness invariants

[achamayou]

@lemmy perhaps I missing something here, but this does not seem like an improvement? It looks like there are fewer distinct states in proportion than before, the frequency of RequestVote has increased even more, and it looks like no successful retirement happened (I notice the run is shorter, but I am not sure that explains it).

[lemmy]

@lemmy perhaps I missing something here, but this does not seem like an improvement? It looks like there are fewer distinct states in proportion than before, the frequency of RequestVote has increased even more, and it looks like no successful retirement happened (I notice the run is shorter, but I am not sure that explains it).

The plots above display the state-space coverage without a de-prioritization of the actions Timeout, ChangeConfiguration, and CheckQuorum (see diff). Meanwhile, the coverage of the property space has improved, as it included the two liveness properties.

With the changes introduced in 47eb0d9, state-space coverage returned to its previous level (see below), while also checking the two liveness properties:

The plots appear more irregular because the data structure used to store the statistics, in this case HyperLogLog, is probabilistic and the duration is less.

[achamayou]

Ah, got it, that's excellent!

[lemmy]

Related: tlaplus/tlaplus#1065