DDX for PostgreSQL — Service Status
Live availability for each public service, plus how fresh the ingested data is. There is no SLA on these services nor guarantee to any schedule or interval; these are observational only.
SERVICE UPTIME
Uptime over the last 15 minutes, 1 hour, and 24 hours. Cells
without enough data yet show —.
| Service | Now | 15 min | 1 hour | 24 hours |
|---|---|---|---|---|
| HTTPS | … | … | … | … |
| MCP | … | … | … | … |
| NNTP | … | … | … | … |
| IMAP | … | … | … | … |
| POP3 | … | … | … | … |
| Blended | … | … | … | … |
Each service is a TCP connect to its local port every 10 s, from the web service on the host: it proves the port accepts connections, not that a request succeeds end to end. The number after the slash is the probe count in the window (15 min holds up to 90, 1 hour up to 360, 24 hours up to 8,640). The history is kept in memory, so it starts again empty when the web service restarts: a window reads — until it has a probe, and the 24-hour window covers only the time since the last restart. Blended is the mean of the five services' percentages, and its count is their probe counts added together. The Now tooltip shows the latency of the latest probe.
Last updated — · source: /healthz
MCP TOOL UTILIZATION
| Tool | Calls | Errors | p50 | p95 |
|---|---|---|---|---|
| Loading… | ||||
Calls per MCP tool (tools/call only) over the selected window: 1m is
30 days, 1yr 365 days. The record starts on 2026-05-20, so 1yr and All are the same
rows until 2027-05-20; calls between 2026-09-07 and 2026-10-02 were not recorded
(the move of MCP to the new server lost the recorder for that time). An error is a
call that returned an error result. p50 and p95 are the server's own time for the
call, not counting the network; the number after the slash is the call count.
Latency colour: green < 0.5 s · yellow 0.5–5 s ·
red > 5 s. Tools with no calls in the window are left out.
Source: /api/v2/tool-stats.
REQUEST GEOGRAPHY (rolling 24h)
Where in the world the last 24 hours of requests came from, summed across every public surface (REST, GraphQL, MCP, git, mailing-list HTML, site). A deeper blue means more requests, on a logarithmic scale; hover a country for the per-API breakdown.
| Country | Total | REST | GraphQL | MCP | Git | Mail HTML | Site |
|---|---|---|---|---|---|---|---|
| Loading… | |||||||
Counted from the web server's access log, one per HTTP request. The MCP
column counts every request to /mcp (session set-up,
tools/list, notifications and streams as well as tool calls), so
it is larger than the tool-call counts above. Left out: /healthz
(load-balancer and monitoring probes), our own hosts, and requests whose
User-Agent is empty or names a crawler, a bot or an HTTP library (curl, wget,
python-requests and the like). Client IPs are looked up in the
DB-IP Lite
IP-to-Country database (CC-BY-4.0); an IP it cannot place counts under
XX. Only (country, API, minute, count) is stored; IP addresses
are not. pg.ddx.io has one origin host, so this is all of its traffic.
Source: /api/v2/geo/24h.
SLOW QUERIES
| Query | Calls | Mean | Max | Rows |
|---|---|---|---|---|
| Loading… | ||||
The 10 database statements with the highest mean time among those that people
caused: web pages, the REST, MCP, GraphQL and UTCP APIs, git over HTTP, and IMAP,
POP3 and NNTP sessions. Counted since agora-web last started
(—). Left out: the service's own work (batch
jobs, scheduled refreshes, cache warming, its health checks and self-probes).
Statements that failed, such as a search cancelled at its time limit, count
as well, and Calls says how many failed. Strings and numbers in the statement
are shown as '?' and N. Max is the slowest single run;
Rows is the total over all calls. Hover the query cell for the statement (cut at
200 characters). Same colour scale as MCP TOOL
UTILIZATION: green < 0.5 s · yellow 0.5–5 s ·
red > 5 s.
HOST HEALTH
CPU, memory, /data disk, disk I/O and network of the host over the last 24 hours, sampled every 60 s and averaged into 5-minute buckets. The figure on the right is the latest sample. A sparkline turns amber when any point in it crosses the warning threshold and red when one crosses critical.
hel1 PRIMARY
CPU is the busy share of all 8 hardware threads. MEM is used memory (total less what the kernel reports available) over total. DISK /data is the filesystem's use. DISK I/O is reads plus writes and NET is received plus sent, in MB/s (MiB). Thresholds: CPU warn 60% / crit 85%, MEM warn 70% / crit 90%, DISK /data warn 85% / crit 95%, DISK I/O warn 20 MB/s / crit 80 MB/s, NET warn 5 MB/s / crit 50 MB/s. CPU, MEM and DISK use a 0–100% scale; DISK I/O and NET scale to the window's own range. hel1 is one dedicated Hetzner server (Intel i7-7700, 4 cores / 8 threads, 64 GB, GeForce GTX 1080; one NVMe disk, two SATA disks mirrored) running PostgreSQL, the web service and the embedding model on the GPU. It is the only host: no failover. Source: /api/v2/host-metrics.
DATA FRESHNESS
How recently each source produced new content. There is no SLA on these intervals; they are observational. The tables refresh every 10 minutes (Other Sources every 30 s).
Mailing Lists and Issue Trackers
| Inbox | Items | Latest message | Avg time to ingest | Next poll |
|---|---|---|---|---|
| Loading… | ||||
Mailing lists, and the forge inboxes that hold a project's GitHub issues and pull
requests, its Trac tickets (postgis-trac) or its gitea pull requests
(postgis-gitea) as threads. Items counts
the messages the list received: a message sent to several lists counts on each of them
(search shows it once), and the patch files an old importer split out of the messages
that carried them do not count. Latest message is the
date of the newest message (its top Received: header, else its
Date:). Avg time to ingest is the mean time from that date to
when the message was written here, over messages of the last 30 days, with the
median in parentheses (/api/v2/ingest-lag). It
leaves out back-fill (a message more than 7 days old when it arrived, and
everything written in the same hour as one) and messages with no readable date;
“-” means no live message in 30 days. Next poll, for the
postgresql.org lists: the probability that the list has mail we have not fetched since
its last poll, and when it is polled next. A list is polled when that probability
reaches 2 %, at least within its own longest gap (one day for a quiet list), at most
every 20 minutes; the five busy lists (pgsql-hackers, -bugs, -general, -committers,
-performance) every 20 minutes, and every list every 6 hours. The tooltip has the last
poll's requests and what it found. postgis-gitea is read slowly on
purpose (one request every few seconds, the first pass at night, UTC), so its
first copy takes about two nights to fill in.
Git Repositories
| Repository | Commits | Symbols | Latest commit | Last fetched |
|---|---|---|---|---|
| Loading… | ||||
Commits counts the commits stored for the repository; Latest commit is the head of
its default branch as of the last fetch, and Symbols the code symbols indexed from
it. For postgres the count is master plus the history of
the live release branches (REL_*_STABLE), tracked and ingested on every
sync — backpatches land within the fetch interval. A freshness check alerts if a
tracked branch falls behind its upstream tip. postgresql-cfbot, the commitfest bot's mirror, keeps only
its branch heads (one cf/<entry> branch per commitfest entry):
its Commits cell shows that head count. The historical University POSTGRES and
University INGRES source releases are not listed here: they never change, and the
archive lists them.
Other Sources
| Source | Kind | Last fetched | Next sync |
|---|---|---|---|
| Loading… | |||
Last fetched is when the source's last sync finished; Next sync is when the next one is due. The documentation is re-read daily, the wiki hourly and the PostgreSQL Hacking Discord server every 5 minutes (Discord is kept for linking people across sources; its messages are not served).
HOW WELL SEARCH WORKS
We measure our own mail and code search against answers that neither this
site nor its operators wrote, and publish the numbers here. Last run
2026-11-02 (04:42–05:40 UTC), against the live service,
build d40fbbf of the agora repository (deployed 2026-11-02, after
release pgddx-2026.11.01).
Method. 314 queries, each with gold Message-IDs taken from public records:
- 170 postgres commits sampled across 2017–2026: the query is the commit
subject, the gold is the thread its
Discussion:trailer links; - 85 commitfest entries (17 each of Committed, Needs review, Rejected, Waiting on Author, Withdrawn): the query is the entry title, the gold is its linked threads;
- 34 plain-language questions written from commit messages, gold from the
Discussion:trailer; - 25 questions from a published outside question set that cite their sources.
A query counts as a hit when the gold thread is among the first k threads returned, so the right thread with a different reply still counts. 313 of the 314 queries could be scored (the gold message of one is not archived here). MRR is the mean of 1/rank of the first gold thread (0 when it is not in the top 10). Latency is the wall time of one MCP tool call from outside, p50 over all 314.
| Tool | hit@1 | hit@10 | MRR | p50 |
|---|---|---|---|---|
hybrid_search | 76.0% | 88.2% | 0.804 | 0.53 s |
retrieve_context | 75.4% | 87.9% | 0.799 | 0.67 s |
By query set (hybrid_search; retrieve_context ranks the
same threads): commit subjects 76.3% hit@1 / 85.2% hit@10; commitfest titles
82.4% / 95.3%; written questions 61.8% / 85.3%; outside questions 72.0% / 88.0%.
Since the run of 2026-11-01 (72.8% hit@1, 87.2% hit@10), a cross-posted discussion
is shown on one list and the semantic search reads twice as many index cells: 4
queries are found now (each at rank 1) and 1 is missed. Against 2026-10-23
(release pgddx-2026.10.23; 70.9% hit@1, 89.5% hit@10), 8 queries found
then are missed now and 4 missed then are found now.
Code search. For 21 literal and regular-expression patterns,
the set of files code_search returns equals the set
git grep -l returns in a postgres clone, at master and at
REL9_6_STABLE: 21 of 21 at each (2026-11-02, against a clone of
master at the commit the archive had indexed, p50 0.25 s).
The harness, the gold set and the scoring are public:
scripts/search-eval
in the agora repository. Re-run it with search-eval.py mail and
search-eval.py code.
INFRASTRUCTURE COSTS
What it costs, roughly, to keep pg.ddx.io running, in USD. The figures come
from infra/docs/cost.txt in the
infra repository
(last reconciled against bills on 2026-07-15) and the provider's published prices;
the last column says which. One dedicated server at Hetzner (hel1)
serves everything; there is no second host, so a host failure means a restore
from the backups in Cloudflare R2 (a continuous WAL archive, a daily dump and a
weekly base backup).
| Provider | Service | Monthly | Annual | Basis |
|---|---|---|---|---|
| Hetzner | hel1 dedicated server (i7-7700 / 64 GB / GTX 1080), the only host | ~$59 | ~$708 | €54/month in cost.txt, at €1 = $1.10; not checked against an invoice here |
| Cloudflare | DNS and CDN (Free plan) | $0 | $0 | Free plan |
| Cloudflare | Load Balancer (two pools, both pointing at hel1) | ~$5.50 | ~$66 | cost.txt; the load balancers are declared in infra/terranix/cloudflare-ddxio.nix; the price is not checked here |
| Cloudflare | R2 storage (backups and archives) | $9–$25 | $108–$300 | $0.015/GB-month (Cloudflare price list). cost.txt estimates ~600 GB ($9); the backup code's own estimate for 30 daily dumps alone is ~1.4 TB (~$21). The bucket size is not checked here |
| Cloudflare | R2 operations (writes and reads) | ~$0.24 | ~$2.88 | cost.txt estimate (~50 k writes, ~10 k reads a month); the first 1 M writes and 10 M reads a month are free, so likely $0 |
| Cloudflare | R2 egress | $0 | $0 | no egress charge (Cloudflare price list) |
| Registrar | ddx.io domain | ~$1 | ~$12 | cost.txt |
| Total | ~$75–$91 | ~$897–$1,089 | ||
| Donations | $0 | $0 | the site does not accept donations | |
| Paid by the operator | ~$75–$91 | ~$897–$1,089 |
Not in the table: outbound mail. The site sends no mail to visitors; operator
alerts go out through a Fastmail relay (infra/modules/mail-forward.nix),
and whether that relay has a working account, and what it costs, is not checked
here. The AWS SES line of earlier versions is gone: SES is not configured anywhere.
Monitoring uses free tiers, and the source repositories on Codeberg cost nothing.
Exoscale, which hosted the service until July 2026, has been closed since
2026-07-15. Nothing on this page reads a
bill: when a provider's invoice and a row disagree, the invoice is right and
infra/docs/cost.txt is the place to fix it.