Both pods on 1.5.0, zero restarts in 35h (cognee-565499f776-bb2g6 up
2026-08-16T18:38:23Z, cognee-mcp-599df8d986-q9gwr up 18:53:50Z). The restart
fix has now held across two container upgrades.
/health answers in 31 ms. One Readiness probe failed: context deadline exceeded event ~48 min before the sweep, no kill — liveness is TCP, so a slow
/health no longer ends the pod.
Resource use: cognee 8m CPU / 1,351 Mi, MCP 4m / 703 Mi.
Rated actuals, 2026-08-09 .. 08-18 UTC:
| Day | Prompt | Completion | Cached | Hit | Cost |
|---|---|---|---|---|---|
| 08-09 | 5,369,061 | 4,059,682 | 174,362 | 3.2% | $1.88 |
| 08-10 | 1,406,172 | 782,337 | 105,489 | 7.5% | $0.41 |
| 08-11 | 2,545,315 | 2,411,892 | 22,563 | 0.9% | $1.03 |
| 08-12 | 1,237,584 | 990,980 | 26,358 | 2.1% | $0.45 |
| 08-13 | 4,508,136 | 3,347,563 | 110,543 | 2.5% | $1.56 |
| 08-14 | 3,016,774 | 2,548,473 | 43,485 | 1.4% | $1.14 |
| 08-15 | 0 | 0 | 0 | — | $0.00 |
| 08-16 | 1,461,324 | 735,871 | 86,790 | 5.9% | $0.40 |
| 08-17 | 7,511,597 | 6,234,932 | 328,781 | 4.4% | $2.77 |
| 08-18 (06:18 UTC) | 1,120,170 | 1,143,921 | 26,294 | 2.3% | $0.48 |
| Total | 28,176,133 | 22,255,651 | 924,665 | 3.3% | $10.11 |
$10.11 over 10 days is $1.01/day, identical to the 08-11 reading. 1.5.0 is cost-neutral. Excluding the zero day it is $1.12/day over nine.
The model name in the billing rows switches from deepseek-v4-flash to
deepseek-v4-flash-0731 at 08-16, which is 30927be landing — a rename in the
bill, not a price change.
Prompt-cache hit rate 3.3% overall, in the same 1–7% band as the 3.5% measured over the clean window on 08-11. PR #85 remains negative; nothing about 1.5.0 moved it.
Two days want explaining:
- 08-15 billed exactly $0.00 with no usage row at all, yet
pipeline_runsrecords 170 runs that day. Either those runs did no LLM work, or something stopped calling out. Unresolved — noted, not investigated. - 08-17 was a $2.77 / 13.7M-token outlier, more than double any other day in the window and the first full day on 1.5.0. Whether that is a one-off re-ingest after the upgrade or a new steady rate needs another few days of data before it means anything.
Corpus went 2,266,251 → 2,712,700 source tokens (+446,449) between 08-11 and 08-18. Fireworks burned 33,857,325 tokens over 08-12 .. 08-18. That is 76 LLM tokens per source token, statistically identical to the 77× measured on 08-06 and the 74× measured pre-remediation on 07-23.
This answers the open blank in COST.md: the amplification ratio against the fresh, post-1.4.2 denominator is unchanged. The appetite fell in absolute terms when the projection went away, but per token ingested it is exactly what it always was.
| 2026-08-11 | 2026-08-18 | |
|---|---|---|
| Database | 2,508 MB | 3,315 MB (~115 MB/day) |
EdgeType_relationship_name |
274,420 rows / 1,300 MB | 331,773 rows / 1,604 MB |
| EdgeType distinct texts | 274,420 of 274,420 | 331,773 of 331,773 |
pipeline_runs |
149 MB / 10,690 rows | 435 MB / 14,479 rows |
run_info size |
avg 13 KB, max 331 KB | avg 29 KB, max 396 KB |
| Graph | 63,317 nodes / 297,392 edges | 75,171 / 359,300 |
| Corpus | 8,895 docs / 2.27M tokens | 10,648 docs / 2.71M tokens |
- Per-edge EdgeType embeddings unchanged on 1.5.0. Still exactly 1:1
distinct — 331,773 rows, 331,773 distinct texts. 1,604 MB, 48% of the
database. The
outbox/draft stands. - Still no ANN index on any of the seven vector tables; all carry only their btree pkey.
pipeline_runsgrowth accelerated to ~41 MB/day from the ~33 MB/day estimated on 08-11, becauserun_infoaverage doubled from 13 to 29 KB. Its two indexes still have zero scans.- The database has passed its pre-maintenance size — 3,315 MB against the 2,664 MB that prompted the 08-06 maintenance window — twelve days after the prune, roughly as predicted.
New this sweep: idx_edge_target_cover on graph_edge is 61 MB with
idx_scan=0 — the single largest never-used index in the database, and the
biggest entry in graph_edge's 205 MB of indexes.