Persistent Cache with Gap-Filling (v2.8.0)

Experiment Overview

Item	Details
Date	2026-01-01
Goal	Eliminate redundant downloads of historical data by removing time-based cache expiry
Environment	alpaca_trading/data/ modules
Status	Success

Context

User noticed that re-running the training notebook caused complete re-downloads of historical data even though:

•Data was downloaded earlier the same day
•Historical data is immutable (past candles never change)
•Only new bars since the last download were needed

The root cause was time-based cache expiry:

•SQLite cache (cache.py): 12-hour TTL via PERSISTED_TTL_HOURS
•Pickle cache (caching_fetcher.py): 3-7 day expiry via cache_expiry_days

v2.8.0 Solution: Persistent Cache + Gap-Filling

Core Principle

Historical market data is immutable. Once downloaded and validated, it should persist indefinitely. Only fetch new bars to fill the gap between cache end and current time.

Changes Made

1. cache.py - SQLite Cache

python

# Before: TTL always checked
def get(self, ..., ttl_hours: int = 24):
    if created_at < ttl_cutoff or expires_at < now_ts:
        self._remove_entry(cache_key)
        return None

# After: TTL is optional (None = no expiry)
def get(self, ..., ttl_hours: Optional[int] = None):
    # Only check TTL if explicitly specified
    if ttl_hours is not None:
        # ... TTL check
    # Otherwise return cached data regardless of age

2. fetcher.py - DataFetcher

python

# Removed
PERSISTED_TTL_HOURS = 12
self._cache_ttl_hours = PERSISTED_TTL_HOURS

# Updated _load_persisted - no TTL check
def _load_persisted(self, symbol: str, timeframe: str) -> pd.DataFrame:
    # No TTL - historical data is immutable
    cached = self._cache.get(symbol, timeframe, start="", end="")
    return cached

3. caching_fetcher.py - CachingDataFetcher

python

class CachingDataFetcher:
    def get_bars(self, symbol, timeframe, lookback_days, **kwargs):
        cached_df = load_from_cache(symbol, timeframe, cache_dir=self._cache_dir)

        if cached_df is not None:
            cache_end = cached_df.index.max()

            # Check if cache covers requested range
            if cache_start <= start_dt and cache_end >= end_dt - tolerance:
                return cached_df  # Complete - no API call

            # Gap-fill: only fetch new bars
            fetch_start = cache_end + timedelta(hours=1)
            new_df = self._fetcher.get_bars(symbol, start=fetch_start, ...)

            # Merge and save
            combined = pd.concat([cached_df, new_df])
            save_to_cache(symbol, combined, ...)
            return combined

Behavior Comparison

Before (Time-Based Expiry)

code

Run 1 (10:00 AM): Fetch 4 years of data [API] -> Cache (12h TTL)
Run 2 (10:30 AM): Cache valid -> [CACHE] instant
Run 3 (11:00 PM): Cache expired -> [API] Fetch 4 years AGAIN

After (Persistent + Gap-Fill)

code

Run 1 (10:00 AM): Fetch 4 years of data [API] -> Cache (persistent)
Run 2 (10:30 AM): Cache complete -> [CACHE] instant
Run 3 (11:00 PM): Cache + gap-fill -> [GAP-FILL] Fetch 13 new bars only

Output Messages

Message	Meaning
`[CACHE] AAPL: 35,040 bars (complete)`	Cache covers full range, no API call
`[GAP-FILL] AAPL: Fetching 2026-01-01 to 2026-01-01...`	Fetching only new bars
`[UPDATED] AAPL: 35,038 + 2 = 35,040 bars`	Merged new bars with cache
`[API] AAPL: Fetching 1460 days...`	No cache, full download

Cache Statistics

New gap_fills counter added:

python

stats = fetcher.get_cache_stats()
# {
#   'cache_hits': 8,      # Returned cached data unchanged
#   'cache_misses': 2,    # No cache, full download
#   'gap_fills': 5,       # Merged new bars with cache
#   'hit_rate': 0.87      # (hits + gap_fills) / total
# }

Failed Attempts

Approach	Result	Why It Failed
Increase TTL to 30 days	Worked but fragile	Still expires eventually, arbitrary cutoff
Check file modification time	Partial	Doesn't verify data completeness

Key Insights

•Historical data is immutable - Past candles never change, so there's no reason to re-fetch them
•Only the edge needs updating - New bars appear at the end of the series
•Time-based expiry is wrong model - For mutable data (news, weather) TTL makes sense; for historical OHLCV it's waste
•Completeness > freshness - Check if cache covers the requested date range, not how old the file is

Files Modified

code

alpaca_trading/data/cache.py:
  - get(): ttl_hours now Optional[int] = None (no expiry by default)

alpaca_trading/data/fetcher.py:
  - Removed PERSISTED_TTL_HOURS constant
  - _load_persisted(): No TTL check
  - _save_persisted(): Uses 10-year TTL (effectively infinite)

alpaca_trading/data/caching_fetcher.py:
  - DEFAULT_*_CACHE_EXPIRY_DAYS = None (no expiry)
  - is_cache_valid(): Just checks file exists
  - get_bars(): Gap-filling logic added
  - get_cache_stats(): Added gap_fills counter

Backward Compatibility

•Existing .pkl cache files work unchanged
•cache_expiry_days parameter still accepted but ignored
•Old caches are automatically upgraded (no migration needed)

References

•Skill: selection-data-caching - Original caching implementation (v2.5.1)
•Skill: data-source-priority - Data fetching hierarchy
•alpaca_trading/data/caching_fetcher.py: Gap-filling implementation