Postgres, storage and a console your team can read in Arabic — running in Egypt.
Pyrabase is a backend platform on the Supabase shape: every project gets its own Postgres, its own
object storage and its own API, with the data resident in Egypt and a console built Arabic-first. The pausing
discipline is what makes the free tier affordable — an idle project releases its compute and keeps its data.
Footprint per awake free project
0.33 GB
was 1.0 GB before D62
RAM needed at launch
74.8 GB
was 105 GB
Free-tier COGS, five years
265,708
USD, after D62 — was 320,098
Infra cost per awake free project
0.7282
USD, −48 % from 1.3947
Why the free tier is affordable, and where that figure comes from
The pack charges only the free projects that are awake in a given month — 45 % still active at year-end,
averaged at a 0.6 factor, so 27 % of a year's free projects are billed in any month. v0.5.6's investor
sentence quoted “$1.1M, the largest unpaid line”; the workbook computes USD 320,098
before D62 and USD 265,708 after. Any material using the old sentence is quoting a
number the model does not produce — that is open item 64, and it is why this page quotes the workbook.
Free-tier COGS over five years (USD)
before D62320,098
after D62265,708
Saved USD 54,390 — smaller than the 60 % v0.5.6 projected, and it is the real number.
Where the footprint came from
1.0 GBbefore
0.33 GBbooked
0.13 GBmeasured
0.13 GB is measured composition and is deliberately unbooked — upside, not a claim.
What is proven, by running it — with the run IDs.
Every claim below is an assertion that fails a CI run, not a design note. Where something is not yet
measured, it says so on the same line.
Six questions, answered on real hardware
- ✓Hibernation releases compute — no pods remain, measured in seconds, not asserted from a spec
- ✓Wake ends on a working database — first successful
SELECT 1 through the Service, not “pod Running”: 12.43 s against a 60 s gate (the spike measured 13.25 s independently)
- ✓External Postgres operator drives a cluster in its own namespace — the arrangement the platform needs
- ✓Footprint measured: 57 MiB Postgres idle, empty database — the other 76 MiB of the composition are estimates and are labelled as such
- ✓Machine printed with every number — 2 cpus, x86_64, because a timing without its machine is not reusable
M0.5 GO run 35219815318 · six of six questions
Tenant storage: five proofs, on a cluster
- ✓A tenant writes to and reads back its own bucket
- ✓Another tenant is denied — and the denial is meaningful, because the allowance held first. A “denied” that appears when everything is denied is not isolation, and the test now fails if it sees that
- ✓The credential secret carries the field names the file service actually reads
- ✓The bucket records an owner — an ownerless bucket is open to any caller
- ✓Deleting a project leaves no identity, credentials, policy or binding
PASSED run 35229388208 · five proofs
Pause and wake: six assertions
- ✓Waking a project that was never paused is harmless — not a one-way door
- ✓After a pause: no pods remain (34.46 s) and the state reads
dormant
- ✓Pausing is not deleting — the cluster, credentials and objects all survive
- ✓The wake ends on a working database inside the gate, and the marker row survives the round trip — the assertion that separates “compute released” from “data lost”
- ✓Components return after the database answers, so nothing dials a dead Service
PASSED run 35232344750 · six of six