One immutable value, compared by identity

A consumer running per audio buffer or per frame needs the current setting on every pass, and needs to know cheaply whether it changed since last time. Two @Volatile fields plus a numeric diff gives you the worst of both: N comparisons on the hot path, and a window where the consumer sees the new value of one field beside the old value of the other.

Bundle the setting instead, hand the consumer a supplier, and compare the reference:

/** One setting, as a single immutable value. */
data class EqualizerCurve(val bandsDb: List<Float>, val preampDb: Float) {
    val isFlat: Boolean = preampDb == 0f && bandsDb.all { it == 0f }
    companion object { val FLAT = EqualizerCurve(emptyList(), 0f) }
}

class Consumer(private val setting: () -> EqualizerCurve) {
    private var applied: EqualizerCurve? = null

    private fun sync() {
        val next = setting()
        if (next === applied) return     // identity, never equality
        applied = next
        rebuild(next)                    // the expensive part: coefficients, allocation, …
    }
}

The producer side is one @Volatile field and a fresh value per write:

@Volatile private var current: EqualizerCurve = EqualizerCurve.FLAT

override fun setEqualizer(bandsDb: List<Float>, preampDb: Float) {
    current = EqualizerCurve(bandsDb, preampDb)   // a NEW object, deliberately
}

Traps

Allocating a fresh value on every write is the point, not waste. It is tempting to skip the allocation when the fields did not change. Don't: the whole scheme rests on "different object ⇒ maybe different setting, same object ⇒ definitely unchanged". Reusing an instance for a value the producer believes is equal reintroduces the diff you were avoiding, one level up, and puts it in the place least able to see the consumer's state. One small object per user gesture is nothing next to N comparisons per buffer for the whole of playback.

=== and == are not interchangeable here, and data class makes the mistake compile. A data class generates equals, so next == applied type-checks and looks more correct — and it costs a list walk on every buffer, which is the exact expense the design exists to avoid. Worse, it makes the rebuild path depend on value equality, so a producer that reuses instances and a consumer that compares by value can each look right in isolation. Write === and say why in a comment beside it.

Identity comparison can only over-rebuild, never under-rebuild — check that direction. A new object with identical contents triggers one needless rebuild (harmless). The dangerous direction is the same object with mutated contents, which the consumer will never notice. That is why the value must be genuinely immutable: List<Float> here is a read-only interface over a list the producer could still hold a mutable reference to. Build it from a defensive copy at the boundary, or accept the list only from callers that construct it fresh.

Bundling is what removes the half-updated read, and that bug is invisible in testing. Two separate volatile fields written in sequence let a consumer between the writes apply a new band set with the old trim — the boost without the headroom that makes room for it. It is one buffer long, it needs a user gesture at exactly the wrong moment, and it never reproduces on demand. One reference write is atomic; there is no "between".

A "nothing to do" flag on the value beats recomputing it per pass. isFlat is computed once in the constructor, so the per-buffer path reads a boolean rather than walking the bands. Compute it as a val in the body — a get() looks the same at the call site and re-walks the list every time it is read. The deciding factor is read frequency, not syntax: on a static table read at composition a get() is fine, and preset-identity-read-back-from-value recommends exactly that.

Invalidate the identity whenever anything else the derived state depends on changes. The coefficients are a function of the setting and the stream format, so after a format change the same object is no longer "applied" — and identity comparison will happily skip the rebuild. Clear the marker wherever the other inputs are negotiated:

override fun onConfigure(format: AudioFormat): AudioFormat {
    …
    applied = null      // forces a rebuild against the new sample rate on the first buffer
    return format
}

Missing this is the subtlest failure in the design, because it needs a second stream at a different sample rate to appear: everything is right until a track arrives at 44.1 kHz after one at 48.

The consumer must sample per pass, not capture at configure time. The supplier only helps if it is invoked on the hot path. Capturing setting() once at stream start is a different, older bug wearing this design's clothes: the setting reaches the output on the next stream and the user reports "it only works from the next track".

One consumer instance per stream, one setting for all of them. The state that forbids sharing a consumer is its own per-stream state, not the setting. Give every player its own consumer and let them all close over the same supplier — a single producer write then covers both halves of a crossfade and every precached player, without anything having to enumerate them. On an engine where that is not true, see sampled-supplier-vs-per-handle-reapply.

Verifying it

# the consumer must compare by identity — expect no hits. The `[^=!]` matters twice over: it skips
# a correct `===` AND a correct `!==`, which is the natural way to write the early return inverted;
# anchoring on `applied` as an operand also skips unrelated equality like `applied?.isFlat == true`.
grep -rnE --include='*.kt' '[^=!]== *applied|applied *==[^=]' .
# and the producer must allocate per write, not reuse
grep -rn --include='*.kt' -A2 'override fun setEqualizer' .

Then two behavioural checks, both of which pass on a broken version if you only run the first:

• write the same values twice and assert rebuild ran twice — that proves identity, not equality, is what is being compared;
• write a value, run one pass, mutate nothing, run a thousand more, and assert rebuild ran once.