Documentation
¶
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type Config ¶
type Config struct {
EvictionThreshold time.Duration // How long a key can be unused before being evicted. (e.g., 30 * time.Minute)
CleanupInterval time.Duration // How often the planner checks for stale keys. (e.g., 5 * time.Minute)
}
Config holds configuration for the planner.
type KeyPlan ¶
type KeyPlan struct {
// contains filtered or unexported fields
}
KeyPlan manages the statistics for a single key and makes decisions about its resolvers. This struct is now entirely lock-free, using a sync.Map to manage its stats.
func (*KeyPlan) SelectStrategy ¶ added in v1.10.2
func (kp *KeyPlan) SelectStrategy(resolvers map[string]*KeyPlanStrategy) *KeyPlanStrategy
SelectStrategy implements the Thompson Sampling decision rule.
func (*KeyPlan) UpdateStats ¶
func (kp *KeyPlan) UpdateStats(plan *KeyPlanStrategy, duration time.Duration)
UpdateStats performs the Bayesian update for the given resolver's statistics.
type KeyPlanStrategy ¶ added in v1.10.2
type KeyPlanStrategy struct {
Type string
/*
In Thompson Sampling, the algorithm models its belief about each strategy's performance as a probability distribution (a "belief curve").
The InitialGuess is the value that the algorithm assumes is the most likely average latency for a strategy it knows nothing about.
To choose a strategy, the algorithm doesn't just use the average; it takes a random sample (a "draw") from each strategy's full belief curve.
Before any data is collected, these draws will be statistically centered around the InitialGuess.
A strategy with a low InitialGuess will have its draws centered around a low number, making it more likely to be chosen initially.
Once a strategy is chosen and its actual performance is measured, that new data point is used to update the belief curve via a Bayesian update.
If a strategy with an InitialGuess of 50ms consistently performs at 10ms, the curve's center will quickly "move" from 50ms down to 10ms. This is how the algorithm learns and adapts.
*/
InitialGuess time.Duration
/*
Lambda represents our confidence in the InitialGuess. How many good runs do we believe we've effectively seen already?
A low Lambda (e.g., 1) means we have very little confidence. The model is "open-minded" and will quickly change its beliefs based on the first few real results. This encourages exploration.
A medium lambda: 5, means a modest trust.
A high Lambda (e.g., 10) means we are very confident. The model is "stubborn" and will require a lot of contradictory data to move away from its initial guess. This encourages exploitation.
*/
Lambda float64
Alpha float64
Beta float64
}
type Planner ¶
type Planner struct {
// contains filtered or unexported fields
}
Planner is the top-level entry point for creating and managing plans for different keys. It is safe for concurrent use and includes a background routine to evict old keys.
func New ¶
New creates a new Planner with the specified configuration and starts its cleanup routine.
func NewNoopPlanner ¶
func NewNoopPlanner() *Planner
func (*Planner) GetKeyPlan ¶
GetKeyPlan retrieves the plan for a specific key, creating it if it doesn't exist.
func (*Planner) StopCleanup ¶
func (p *Planner) StopCleanup()
StopCleanup gracefully terminates the background cleanup goroutine.
type ThompsonStats ¶
type ThompsonStats struct {
// contains filtered or unexported fields
}
ThompsonStats holds the parameters for the Normal-gamma distribution, which models our belief about the performance (execution time) of a strategy.
func NewThompsonStats ¶
func NewThompsonStats(initialGuess time.Duration, lambda, alpha, beta float64) *ThompsonStats
NewThompsonStats creates a new stats object with a diffuse prior, representing our initial uncertainty about a strategy's performance.
func (*ThompsonStats) Sample ¶
func (ts *ThompsonStats) Sample(r *rand.Rand) float64
Sample draws a random execution time from the learned distribution. This is the core of Thompson Sampling: we sample from our belief and act greedily on that sample.
func (*ThompsonStats) Update ¶
func (ts *ThompsonStats) Update(duration time.Duration)
Update performs a Bayesian update on the distribution's parameters using the new data point (the observed execution duration). It is the responsibility of the caller to enforce synchronization if multiple goroutines may call Update concurrently.
Source Files