Merge #19968: doc: clarify CRollingBloomFilter size estimate

d9141a0002bb508b2e94e206a1bd28ef8f97ffde doc: clarify CRollingBloomFilter size estimate (Anthony Towns)

Pull request description:

  Based on #19130, this change improves the comment for `CRollingBloomFilter` in `bloom.h`:

  - Give examples to illustrate the heuristic "1.8 bytes per element per factor 0.1 of false positive rate"
  - Add some Python code which can be copy/pasted for convenient filter size calculation (in an interpreter)
  - Reconcile the newly added code with the existing approximation

ACKs for top commit:
  laanwj:
    ACK d9141a0002bb508b2e94e206a1bd28ef8f97ffde

Tree-SHA512: e7138b3c531883a750ead06368975c750863fde7ef6f2633b137eca011079226e9205316217322014399fba05a48f294c788dd700bb7d479c58fe1f23e40419f

1 files changed, 12 insertions, 1 deletions

diff --git a/src/bloom.h b/src/bloom.h
index 9307257852..24dc607cd9 100644
--- a/src/bloom.h
+++ b/src/bloom.h
@@ -94,7 +94,18 @@ public:
  * insert()'ed ... but may also return true for items that were not inserted.
  *
  * It needs around 1.8 bytes per element per factor 0.1 of false positive rate.
- * (More accurately: 3/(log(256)*log(2)) * log(1/fpRate) * nElements bytes)
+ * For example, if we want 1000 elements, we'd need:
+ * - ~1800 bytes for a false positive rate of 0.1
+ * - ~3600 bytes for a false positive rate of 0.01
+ * - ~5400 bytes for a false positive rate of 0.001
+ *
+ * If we make these simplifying assumptions:
+ * - logFpRate / log(0.5) doesn't get rounded or clamped in the nHashFuncs calculation
+ * - nElements is even, so that nEntriesPerGeneration == nElements / 2
+ *
+ * Then we get a more accurate estimate for filter bytes:
+ *
+ *     3/(log(256)*log(2)) * log(1/fpRate) * nElements
  */
 class CRollingBloomFilter
 {