feat(deferred): Improve default retry strategy

[xingzihai]

Summary

This PR improves the default retry strategy for Rage::Deferred tasks as described in #251.

Changes

1. Increased MAX_ATTEMPTS from 5 to 15

   • With 5 retries, tasks exhausted all attempts in ~2.7 minutes
   • With 15 retries, tasks have ~2 days before exhausting all attempts
   • This gives developers more time to discover and fix underlying issues

2. Changed backoff formula from exponential to polynomial

   • Old: rand(5 * 2**attempt) + 1 (exponential growth)
   • New: (attempt**4) + 10 + (rand(15) * attempt) (polynomial growth)
   • This spaces retries further apart with each attempt

Retry timing comparison

Attempt	Old (avg)	New (avg)
1	~3s	~18s
2	~6s	~41s
3	~11s	~2m
4	~21s	~5m
5	~41s	~11m
10	-	~2.8h
15	-	~14.1h

Total retry time

• Old: ~2.7 minutes (5 retries)
• New: ~2 days (15 retries)

Testing

Updated spec tests to verify the new backoff intervals:

• __next_retry_in(0, nil) returns between 11-25 seconds
• __next_retry_in(1, nil) returns between 26-54 seconds
• __next_retry_in(15, nil) returns between 50635-50845 seconds (~14h)
• __next_retry_in(16, nil) returns nil (max exceeded)

Closes #251

[anuj-pal27]

I think there are two changes needed to align with issue #251 and maintainer feedback:

• Set MAX_ATTEMPTS to 20 (not 15).
  In the issue discussion, the direction moved toward 20 to better cover real-world outages (especially weekend/longer incidents). 15 improves things, but 20 gives a much safer retry window by default, which is the main goal of #251.
• Don’t increment attempt in __default_backoff: Using attempt + 1 shifts the retry timing and skips the intended first tier (~18s avg), making retry 1 behave like retry 2.

We should keep the formula exactly as specified:

def __default_backoff(attempt)
  (attempt**4) + 10 + (rand(15) * attempt)
end

@rsamoilov — is this understanding correct?