How do I keep recommendations fresh when inventory changes fast?

To keep recommendations fresh when inventory changes fast, use a system that updates on both a schedule and on events.

1) Trigger updates from inventory events

When stock changes, fire events such as:

• item back in stock
• item sold out
• price changed
• new item added
• item removed or discontinued

Use those events to:

• remove unavailable items from candidate lists
• boost newly available items
• retrain or refresh features tied to inventory state

2) Separate “candidate generation” from “final ranking”

A good pattern is:

• Candidate generation: broad pool of potentially relevant items
• Ranking/filtering: apply live inventory, price, and availability constraints right before showing results

This way, even if your model is a little stale, the final layer keeps recommendations current.

3) Cache smartly, not blindly

Cache:

• static features like category, brand, embeddings
• user preference vectors
• precomputed candidate sets

Avoid caching:

• live stock availability
• current price
• delivery estimates

Use short TTLs for volatile data.

4) Add real-time business rules

Before displaying recommendations, apply rules like:

• exclude out-of-stock items
• prefer items with higher stock
• down-rank items with low remaining inventory
• prioritize items that need to be cleared
• avoid recommending items unlikely to ship quickly

5) Refresh features often

If your model uses inventory-dependent features, update them frequently:

• stock count
• days of supply
• sales velocity
• replenishment ETA
• popularity over last hour/day

For very fast-moving catalogs, use micro-batches or streaming feature updates.

6) Blend long-term preferences with short-term trends

Use:

• long-term user history for relevance
• short-term signals for freshness:
  • recent clicks
  • trending items
  • newly stocked items
  • seasonal demand

This helps recommendations stay current without forgetting user taste.

7) Use fallbacks

If inventory changes too fast:

• show similar substitute items
• recommend available alternatives in the same category
• surface “back in stock” alerts for items users wanted

8) Monitor freshness metrics

Track:

• out-of-stock recommendation rate
• recommendation latency vs inventory updates
• click-through on newly stocked items
• conversion after stock changes
• stale-item exposure rate

These tell you whether freshness is actually improving.

Simple practical setup

A common architecture is:

1. Inventory system emits events
2. Feature store updates volatile attributes
3. Candidate cache refreshes frequently
4. Final ranking service filters by live availability
5. Monitoring alerts if stale items leak through

If you want, I can also give you:

• a reference architecture
• a real-time ML pipeline design
• or sample rules/code for inventory-aware recommendations.