How It Works Pricing Benchmarks
vs Redis Docs Resources Blog
Start Free Trial
New Primitive

Your Cache Proves It's Correct.
Continuously.

Statistical sampling compares cached values against source-of-truth in real-time. Detects poisoning, corruption, and drift. Auto-repairs individual keys. Reports a consistency score per key prefix.

99.99%
Consistency Score
Auto
Repair
<5s
Poisoning Detection
Zero
False Trust
The Problem

Every Cache Trusts Its Own Data Blindly

Your cache serves millions of responses per second. Not one of them is verified. The cached value could be wrong, stale, or maliciously injected — and nobody would know until a customer reports it.

🚫
No Verification, No Sampling, No Proof
Every cache in production today operates on blind trust. A value is written once and served indefinitely until TTL expires or the key is explicitly invalidated. There is no mechanism to verify that the cached value still matches the source. No sampling. No comparison. No proof of correctness. You trust the cache because you have no alternative.
0% of cached values are ever verified
Cache Poisoning Is a Real Attack Vector
An attacker who can write to the cache controls what the application sees. Compromised cache entries serve malicious data to every user, every request, until the TTL expires or someone manually flushes the key. Nobody detects this. There is no monitoring for cache correctness. The poisoned value looks identical to a legitimate one.
Poisoned data served to every user until TTL expires
💨
Silent Drift Nobody Knows About
CDC events get dropped. Race conditions between concurrent writes produce stale snapshots. Partial updates leave half-written objects in cache. Network partitions cause divergence between cache and source. The cache silently drifts from truth, and nobody knows until a customer reports seeing wrong data. By then, the drift has been serving incorrect responses for hours.
Drift is invisible until a customer reports it
How It Works

Sample. Compare. Repair. Report.

A background process continuously samples cached keys, fetches the authoritative value from the source-of-truth, and compares them. Mismatches are auto-repaired. Everything is reported as a consistency score per key prefix.

Self-Healing Consistency — Sampling Loop
Step 1
Sample N Keys
configurable rate/min
Step 2
Fetch Source
DB / API / upstream
Step 3
Compare
cached vs authoritative
Match
Score ++
consistency confirmed
↓ mismatch
Mismatch Detected
Auto-Repair
re-fetch & update key
Log
Divergence
key / expected / actual
Alert
Security Team
if threshold breached
Consistency Score Formula
(sampled_matches / total_samples) × 100
Calculated per key prefix. Exposed on the dashboard in real time.
# Enable self-healing consistency for a key prefix CONSISTENCY user:* SAMPLE_RATE 1000/min SOURCE postgres://db/users # The cache now samples 1000 user:* keys per minute # For each sample: fetch from Postgres, compare, repair if mismatched # Check consistency score on the dashboard or via API GET /api/consistency/score # { # "user:*": { "score": 99.997, "samples": 14832, "repairs": 0 }, # "product:*": { "score": 99.981, "samples": 14519, "repairs": 3 }, # "session:*": { "score": 99.200, "samples": 14201, "repairs": 112 } <-- investigate # } # Divergence event log (auto-generated) # { # "timestamp": "2026-03-27T14:32:01.847Z", # "key": "product:8821:price", # "expected": "{\"amount\":29.99,\"currency\":\"USD\"}", # "actual": "{\"amount\":0.01,\"currency\":\"USD\"}", <-- poisoned! # "action": "auto-repaired", # "source": "postgres://db/products" # }

Background Process, Zero Read-Path Impact

Self-healing consistency runs as a background sampling loop. It does not intercept reads. It does not add latency to writes. It does not sit in the hot path. Your GET and SET operations are completely unaffected. The sampling rate is configurable per key prefix — sample 100 keys per minute for low-traffic prefixes, 10,000 per minute for critical ones.

Repair is surgical. When a mismatch is detected, only the diverged key is re-fetched and updated. There is no full invalidation, no cache flush, no thundering herd. One key is wrong, one key is fixed.

Statistical Confidence at Scale

Sampling 1,000 keys per minute from a namespace of 1 million keys gives you statistically significant coverage. Within an hour, you have sampled 60,000 keys. Within a day, 1.44 million — more than your entire keyspace. Any systematic corruption or drift surfaces quickly through the consistency score.

Random sampling catches both targeted attacks (a specific key is poisoned) and systemic drift (a CDC pipeline is dropping events). The probability of a poisoned key surviving undetected decreases exponentially with each sampling cycle.

Security

The Cache Vendor That Detects Cache Poisoning

Cache poisoning is an attack vector that nobody talks about. An attacker who compromises a cache write path can serve malicious data to every user. Self-healing is the only defense that exists.

🛡
Detection Within Seconds
Statistical sampling catches poisoned keys within seconds of injection. At 1,000 samples per minute, a poisoned key in a namespace of 10,000 keys has a 10% chance of being sampled every minute. Within 5 minutes, the probability of detection exceeds 99.99%. The attacker's window of exploitation is measured in seconds, not hours.
99.99% detection probability within 5 minutes
🔧
Automatic Remediation
When a poisoned key is detected, the cache does not just log it. It auto-repairs: fetches the correct value from the source-of-truth, overwrites the poisoned entry, logs the full divergence event (timestamp, key, expected value, actual value), and alerts the security team. The attack is neutralized without human intervention.
Detect → repair → log → alert in one cycle

No other caching system detects cache poisoning. Redis does not verify its own data. Memcached does not sample keys. Hazelcast does not compare cached values to source. Every cache in production today trusts its own data unconditionally. Self-healing consistency is the first mechanism that treats the cache as an untrusted data store and continuously verifies it against the authoritative source.

For enterprise security teams evaluating caching vendors, this is a differentiated capability. Being the cache vendor that detects and auto-repairs cache poisoning is a pitch that no competitor can match.

Dashboard

Consistency Score Per Prefix

Every key prefix gets its own consistency score. Green means healthy. Yellow means investigate. Red means alert. One number that tells you whether your cache is serving correct data.

user:*
99.99%
product:*
99.98%
session:*
99.90%
pricing:*
98.20%
99.99% = Healthy
Your cache is serving correct data. The consistency score confirms that sampled keys match the source-of-truth. Normal operating state. No action required.
99.9% = Investigate
Divergence detected at a low rate. Could be a CDC pipeline lag, a race condition in a write path, or a transient network issue. Auto-repair is handling it, but the root cause should be investigated.
🚨
<99% = Alert
Significant divergence. Could indicate a systematic issue: a broken CDC connector, a cache poisoning attack, or a misconfigured write path. Security team is alerted. Auto-repair is actively fixing keys while the root cause is triaged.
Divergence Events
Timeline
When, which key, what changed
Auto-Repair Count
Per Hour
Keys repaired without intervention
Sampling Rate
Per Prefix
Configurable samples/minute
Source Latency
P50 / P99
Source-of-truth response time
Use Cases

Who Needs Self-Healing Consistency

Four categories of teams where cache correctness is not optional.

🛡
Security Teams
Cache poisoning detection. The only caching system that continuously verifies cached data against source-of-truth and auto-repairs poisoned entries. Alerts on divergence. Logs every mismatch with full context for forensic analysis.
Poisoning detection + auto-remediation
Platform Teams
Silent drift detection. Catches CDC events that were dropped, race conditions in write paths, and partial updates that left inconsistent data in cache. The consistency score surfaces drift before any customer sees it.
Catch drift before customers do
📋
Compliance Teams
Provable cache accuracy. The consistency score is an auditable metric that proves your cache is serving correct data. For regulated industries (finance, healthcare, government), this is evidence that your data layer meets accuracy requirements.
Auditable proof of data accuracy
🔍
Debugging
When did this key diverge? The divergence event log gives you a timeline of every mismatch: timestamp, key, expected value, actual value, source, and repair action. Turns a "the cache is wrong" ticket into an actionable root-cause investigation.
Full divergence timeline per key
A cache that doesn't verify itself is a cache you can't trust.
Self-healing consistency is the difference between blind trust and proof.
FAQ

Frequently Asked Questions

What is self-healing consistency in caching?

Self-healing consistency is a background process that statistically samples cached keys and compares their values against the source-of-truth (database, API, or upstream cache). When a mismatch is detected, the cache automatically re-fetches the correct value and updates the entry without requiring a full cache invalidation. The process reports a consistency score per key prefix so operators can monitor cache accuracy in real time.

How does Cachee detect cache poisoning?

Cachee's self-healing consistency engine continuously samples cached values and compares them to the authoritative source. If an attacker compromises a cache write path and injects malicious data, the statistical sampling detects the mismatch within seconds. The poisoned key is automatically repaired, the divergence is logged with full context (timestamp, key, expected value, actual value), and an alert is sent to the security team.

What is a cache consistency score?

A consistency score is the ratio of sampled cache keys that match their source-of-truth value, expressed as a percentage. Cachee calculates this per key prefix and exposes it on the dashboard. A score of 99.99% means your cache is healthy. A score of 99.9% warrants investigation. A score below 99% triggers an alert. The consistency score provides continuous, auditable proof that your cache is serving correct data.

Does self-healing consistency add latency to cache reads?

No. Self-healing consistency runs as a background process on a configurable sampling schedule. It does not intercept or add latency to any cache read or write operation. The sampling, comparison, and repair all happen asynchronously. Your GET and SET operations are completely unaffected.

How is self-healing different from TTL-based expiration?

TTL expiration blindly evicts keys after a fixed time period, regardless of whether the cached value is still correct. Self-healing consistency verifies whether the value is actually correct by comparing it to the source-of-truth. Correct values stay cached. Incorrect values are repaired in place. This means you get higher hit rates (no unnecessary evictions) and stronger correctness guarantees (actual verification instead of time-based guessing).

Stop Trusting Your Cache Blindly.
Prove It's Correct. Continuously.

Self-healing consistency detects poisoning, catches drift, auto-repairs corruption, and gives you a consistency score you can put in front of auditors.

Start Free Trial Schedule Demo