Utilization

How well the cache is being used right now: footprint against quota per tenant, and how each instance splits between the hot and cold tiers.

L2 footprint

87.0 GB

live from GET /quota

Of summed quota

66%

66.3 GB of 100.0 GB · 3 tenants with quotas

Tenants

3 with quotas

Tenants over quota

1 near the limit (≥ 85%) · 3 tenants with quotas

Tier balance

L1 (hot) vs L2 per MP server, on one shared scale — L2 from each instance's GET /cache/objects listings, L1 from the coordinator's directory.

mp-server-a11,663
2161,44713% in L1
mp-server-a21,491
1901,30113% in L1
mp-server-b11,050
11393711% in L1

L1519L23,685fleet 12.3% in L1

Top tenants

Usage and quota are live from GET /quota, largest footprint first.

Tenant	Usage	Quota	% of fleet
team-search	29.5 GB	31.0 GB95% used	33.8%
team-support	24.1 GB	57.0 GB42% used	27.7%
default	20.8 GB	no quota	23.9%
batch-eval	12.7 GB	12.0 GB106% over	14.6%
4 tenants	87.0 GB	100.0 GB	100.0%

Trends live elsewhere

Everything on this page is a point-in-time read of the coordinator. Hit rates and history come from the serving layer (Prometheus / vLLM) and the cache-event stream — they arrive with that connection.

Cache hit rate per model and tenant
Usage and eviction history over time
Time-to-first-token saved by a cache hit
Tokens served from cache vs recomputed

The directory can say a chunk is resident on mp-server-a1; only the serving layer knows whether a request reused it.