--- title: Monitoring | Runware Docs url: https://runware.ai/docs/serverless/api-monitoring description: Read an app's events and failed requests, and the named metric and log queries behind them. relatedDocuments: - https://runware.ai/docs/serverless/api - https://runware.ai/docs/serverless/monitoring - https://runware.ai/docs/serverless/api-workers --- ## Introduction Two shapes of read. **Events and failed requests hang off an app** and are listed directly. **Metrics and logs go through named queries**: you ask the catalog what it can answer, then read one query by its id. [Monitoring](https://runware.ai/docs/serverless/monitoring) covers which of these answers which question. This page is the routes. > [!NOTE] > Reading logs, metrics and an app's request errors draws on a read allowance for your organization. When it runs out those routes answer `429` with a `Retry-After`, and a client that ignores it keeps being refused. ## List app events GETapi.serverless.runware.ai/v1/apps/{appId}/events > [!NOTE] > Events are the platform's record of what it did: `deploy`, `scaling`, `audit` and `error`. `scaling` is usually the one to read first when an app costs more or answers slower than you expected, because it says when workers appeared and disappeared. ### [Request](#list-app-events-request) Path ### [appId](#list-app-events-path-appid) - **Type**: `string` - **Required**: true - **Min**: `6` - **Max**: `30` Immutable app identifier, unique among the authenticated organization's live apps. Query ### [limit](#list-app-events-query-limit) - **Type**: `integer` - **Format**: `int32` - **Min**: `1` - **Max**: `100` - **Default**: `20` Maximum number of items to return. ### [cursor](#list-app-events-query-cursor) - **Type**: `string` Opaque pagination cursor returned by a previous call, as `nextCursor` or, on the operations that offer one, `prevCursor`. ### [type](#list-app-events-query-type) - **Type**: `string` **Allowed values**: `deploy` `scaling` `audit` `error` ### [Response](#list-app-events-response) ### [nextCursor](#list-app-events-response-nextcursor) - **Type**: `string (nullable)` Cursor for the next page. Null when there are no more items. ### [data](#list-app-events-response-data) - **Type**: `object[] (7 properties)` #### [id](#list-app-events-response-data-id) - **Type**: `string (uuid)` - **Required**: true - **Format**: `UUID v4` #### [appId](#list-app-events-response-data-appid) - **Type**: `string` - **Required**: true - **Min**: `6` - **Max**: `30` Immutable app identifier. Unique among the authenticated organization's live apps: it cannot be changed after creation, and it becomes available again once the app it named reaches `deleted`. #### [workerId](#list-app-events-response-data-workerid) - **Type**: `string (uuid) (nullable)` - **Format**: `UUID v4` #### [endpointId](#list-app-events-response-data-endpointid) - **Type**: `string (uuid) (nullable)` - **Format**: `UUID v4` #### [type](#list-app-events-response-data-type) - **Type**: `string` - **Required**: true **Possible values**: `deploy` `scaling` `audit` `error` #### [message](#list-app-events-response-data-message) - **Type**: `string` - **Required**: true Human-readable description of what happened. Never blank. #### [createdAt](#list-app-events-response-data-createdat) - **Type**: `string (date-time)` - **Format**: `date-time` ### [Errors](#list-app-events-errors)
StatusWhen
400The request was malformed and could not be parsed (e.g. invalid JSON). A well-formed request that fails validation returns 422 instead.
401Missing or invalid credentials
404Resource not found
422The request was well-formed but semantically invalid (e.g. a missing or out-of-range field). A request that could not be parsed returns 400. The errors array carries one entry per offending field.
500Unexpected server error
503A required service is temporarily unavailable
## List request errors GETapi.serverless.runware.ai/v1/apps/{appId}/errors One page of failed inference requests for this app, newest first. Omit `statusClass` for both 4xx and 5xx. The cursor is opaque and is only valid with the same `window` and `statusClass` it was issued under. > [!NOTE] > Omitting `statusClass` returns `4xx` and `5xx` together, which is usually what you want first: the split between them is the diagnosis. The cursor is only valid with the `window` and `statusClass` it was issued under, so changing either means starting again. ### [Request](#list-app-errors-request) Path ### [appId](#list-app-errors-path-appid) - **Type**: `string` - **Required**: true - **Min**: `6` - **Max**: `30` Immutable app identifier, unique among the authenticated organization's live apps. Query ### [limit](#list-app-errors-query-limit) - **Type**: `integer` - **Format**: `int32` - **Min**: `1` - **Max**: `100` - **Default**: `20` Maximum number of items to return. ### [cursor](#list-app-errors-query-cursor) - **Type**: `string` Opaque pagination cursor returned by a previous call, as `nextCursor` or, on the operations that offer one, `prevCursor`. ### [window](#list-app-errors-query-window) - **Type**: `string` - **Default**: `24h` The range to search. Same closed ladder as the metrics queries. Defaults to the last 24 hours. **Allowed values**: `1h` `6h` `24h` `7d` `30d` ### [statusClass](#list-app-errors-query-statusclass) - **Type**: `string` Narrow to one error class. Omit for both 4xx and 5xx. **Allowed values**: `4xx` `5xx` ### [Response](#list-app-errors-response) ### [nextCursor](#list-app-errors-response-nextcursor) - **Type**: `string (nullable)` Cursor for the next page. Null when there are no more items. ### [data](#list-app-errors-response-data) - **Type**: `object[] (4 properties)` - **Required**: true #### [time](#list-app-errors-response-data-time) - **Type**: `integer` - **Required**: true - **Format**: `int64` #### [level](#list-app-errors-response-data-level) - **Type**: `string` #### [body](#list-app-errors-response-data-body) - **Type**: `string` - **Required**: true #### [fields](#list-app-errors-response-data-fields) - **Type**: `object` ### [Errors](#list-app-errors-errors)
StatusWhen
400The request was malformed and could not be parsed (e.g. invalid JSON). A well-formed request that fails validation returns 422 instead.
401Missing or invalid credentials
404Resource not found
422The request was well-formed but semantically invalid (e.g. a missing or out-of-range field). A request that could not be parsed returns 400. The errors array carries one entry per offending field.
429The organization's read allowance is exhausted. Retry-After says how long to wait. A client that ignores it will keep being refused.
500Unexpected server error
502An upstream service was unreachable or refused the request
503A required service is temporarily unavailable
504An upstream service did not answer in time
## The query catalog Every metric and log read names a query the platform already knows how to answer, and **the catalog is the source of those ids**: discover them from it rather than carrying a list of your own. Each entry also reports the `windows` it can answer. **A window missing from that list has no stored series behind it**, so offering it would render an empty chart. Some queries are deliberately absent from the catalog because they back a specific surface: the `runtime` object on an endpoint, and an app's overview. **Reading one by id still works**, so an absence from the catalog is about what a client should offer rather than what it can ask for. ## List queries GETapi.serverless.runware.ai/v1/metrics/queries The catalog: every named query, its unit and aggregation, the selectors it accepts, the series it returns, and the windows actually backed by stored series. This is the source of query ids for the Metrics and Logs tabs. Clients discover ids here rather than carrying a list of their own, and render window tabs from `windows` rather than from the full ladder, so a window whose storage tier has no backing series stays invisible instead of rendering a tab with nothing behind it. The list sparklines `apps_request_volume` and `endpoints_request_volume` are the only list-page queries here. The other list-page runtime queries and the app overview queries are not charts, so this catalog omits them. `getMetricSeries` serves them by id. ### [Response](#list-insights-queries-response) ### [metrics](#list-insights-queries-response-metrics) - **Type**: `object[] (6 properties)` - **Required**: true #### [id](#list-insights-queries-response-metrics-id) - **Type**: `string` - **Required**: true The value to pass as `queryId`. #### [unit](#list-insights-queries-response-metrics-unit) - **Type**: `string` - **Required**: true The unit of every value in this query's series, e.g. `req/min`. #### [aggregation](#list-insights-queries-response-metrics-aggregation) - **Type**: `string` - **Required**: true How each bucket was reduced, e.g. `avg` or `p50`. Carried so a tooltip can say what a point means rather than presenting a bucket reduction as an instant reading. #### [windows](#list-insights-queries-response-metrics-windows) - **Type**: `string[]` - **Required**: true The windows this query can answer. A window absent here has no stored series behind it, so a client should not offer it. #### [selectors](#list-insights-queries-response-metrics-selectors) - **Type**: `string[]` - **Required**: true The narrowings this query accepts. Anything else is rejected. #### [series](#list-insights-queries-response-metrics-series) - **Type**: `string[]` - **Required**: true The stable series ids this query returns. Empty when those ids are not known until the request (`apps_request_volume` keys each line on an app id). ### [logs](#list-insights-queries-response-logs) - **Type**: `object[] (6 properties)` - **Required**: true #### [id](#list-insights-queries-response-logs-id) - **Type**: `string` - **Required**: true The value to pass as `queryId`. #### [unit](#list-insights-queries-response-logs-unit) - **Type**: `string` - **Required**: true The unit of every value in this query's series, e.g. `req/min`. #### [aggregation](#list-insights-queries-response-logs-aggregation) - **Type**: `string` - **Required**: true How each bucket was reduced, e.g. `avg` or `p50`. Carried so a tooltip can say what a point means rather than presenting a bucket reduction as an instant reading. #### [windows](#list-insights-queries-response-logs-windows) - **Type**: `string[]` - **Required**: true The windows this query can answer. A window absent here has no stored series behind it, so a client should not offer it. #### [selectors](#list-insights-queries-response-logs-selectors) - **Type**: `string[]` - **Required**: true The narrowings this query accepts. Anything else is rejected. #### [series](#list-insights-queries-response-logs-series) - **Type**: `string[]` - **Required**: true The stable series ids this query returns. Empty when those ids are not known until the request (`apps_request_volume` keys each line on an app id). ### [Errors](#list-insights-queries-errors)
StatusWhen
401Missing or invalid credentials
500Unexpected server error
502An upstream service was unreachable or refused the request
503A required service is temporarily unavailable
504An upstream service did not answer in time
## Read a metric query GETapi.serverless.runware.ai/v1/metrics/queries/{queryId}/series Returns one chart's data: a single timestamp axis shared by every series, and one dense value array per series aligned to it. Values are dense and positionally aligned to `t`, with an explicit `null` wherever there was no sample. Each timestamp is the END of its bucket, so `window.to` is inclusive and equals the last timestamp in `t`, while `window.from` is exclusive and is one `step_s` before the first. An organization with no metrics yet is answered with the full axis and all-null series rather than an error. `apps_request_volume` returns one series per app: 96 quarter-hour request counts over `window=24h` (`step_s` 900, unit `requests`). Repeat `appId` once per id on the current list page to pad idle apps with all-null series, in request order. A series is named for the live app behind it, so a reused app id reports its own generation's traffic and not the one before it. The same `appId` pad applies to the list-scoped `apps_error_volume` and `apps_request_duration` queries. Those feed `listApps` / `getApp` `runtime` and are omitted from the catalog. Other queries reject `appId`. Other windows are not available for these queries. `endpoints_request_volume` is the endpoints-list counterpart: 96 quarter-hour request counts per endpoint over `window=24h` (`step_s` 900, unit `requests`). It requires the public app id. Repeat `endpointId` once per id on the current `listEndpoints` page to pad idle endpoints with all-null series, in request order. Buckets that started before that endpoint row's `createdAt` are null, so a removed-then-readded path does not inherit the previous row's traffic. The same app + `endpointId` pad and `createdAt` clip apply to the list-scoped `endpoints_request_rate`, `endpoints_latency_p95`, `endpoints_latency_p99`, `endpoints_error_volume` and `endpoints_request_count` queries (`window=1h` only, 5-minute step). Those feed `listEndpoints` / `getEndpoint` `runtime` and are omitted from the catalog. Other queries reject `endpointId`. Other windows are not available for these queries. `throughput`, `latency` and `error_rate` accept `endpoint` (the allocated UUID) together with the public app id, so the Metrics-tab charts can be scoped to one path. Other cataloged queries reject `endpoint`. `endpoint` is a matcher, not the `endpointId` pad. `gpu_utilisation`, `gpu_power_usage`, `gpu_power_utilisation`, `gpu_temperature` and `gpu_memory_used` require the public app id and accept `worker` (`Worker.id`, the Kubernetes pod UID), so the Metrics tab can chart an app and a worker page can chart one live worker. Other cataloged queries reject `worker`. `gpu_count` stays app-scoped. Three queries serve one app's overview, all over `window=24h` at `step_s` 900 and all requiring the public app id: `app_traffic_24h` returns `requests`, `client_errors` (4xx) and `server_errors` (5xx) as request counts. `app_worker_seconds_24h` returns `startup`, `execution` and `idle` as worker-seconds, whose three values in a bucket sum to that bucket's worker time, and `app_cold_starts_24h` returns `cold_starts` as a count. They report counts and totals rather than rates or ratios, so a per-minute figure is a bucket value divided by `step_s / 60` and a 24h ratio is one summed axis over another, summing first and dividing once, because averaging a per-bucket ratio across the axis does not give the 24h ratio. These queries are absent from `listInsightsQueries`: they back the overview rather than the Metrics tab. Other windows are not available for them. > [!NOTE] > One chart per read: a single timestamp axis shared by every series, and one dense value array per series aligned to it, with an explicit `null` wherever there was no sample. **Each timestamp is the end of its bucket**, so a point at `t[i]` covers the `step_s` seconds before it. An organization with no metrics yet is answered with the full axis and all-null series rather than an error. ### [Request](#get-metric-series-request) Path ### [queryId](#get-metric-series-path-queryid) - **Type**: `string` - **Required**: true A query id from the catalog. Deliberately not an enum: the downstream registry is the source of ids, and an enum here would be a second list to keep in step with it. Query ### [window](#get-metric-series-query-window) - **Type**: `string` - **Required**: true The time window. A closed set rather than a free-form range, because every distinct range defeats the server-side cache alignment that makes a sliding window cheap. Only the windows a query lists in the catalog can be asked of it. **Allowed values**: `1h` `6h` `24h` `7d` `30d` ### [pinnedTo](#get-metric-series-query-pinnedto) - **Type**: `integer` - **Format**: `int64` Fixes the window's inclusive end, the last timestamp in `t`, so that several calls making up one visual share an axis instead of racing the clock between them. Must be aligned to the window's step, no newer than the newest readable edge, and inside retention. ### [deployment](#get-metric-series-query-deployment) - **Type**: `string` - **Min**: `6` - **Max**: `30` Narrow to one app. ### [endpoint](#get-metric-series-query-endpoint) - **Type**: `string (uuid)` - **Format**: `UUID v4` Narrow to one endpoint within an app. The value is the allocated `endpoints.id` (UUID), never the customer-authored path. Accepted by `throughput`, `latency` and `error_rate`, and requires the public app id. A query that does not declare this selector rejects it rather than ignoring it. ### [statusClass](#get-metric-series-query-statusclass) - **Type**: `string` Narrow to one response class. **Allowed values**: `2xx` `4xx` `5xx` ### [region](#get-metric-series-query-region) - **Type**: `string` Narrow to one region. No series carries a region label yet, so no query currently accepts this and supplying it is rejected rather than ignored. ### [worker](#get-metric-series-query-worker) - **Type**: `string (uuid)` - **Format**: `UUID v4` Narrow to one live worker within an app. The value is `Worker.id` (the Kubernetes pod UID). Accepted by `gpu_utilisation`, `gpu_power_usage`, `gpu_power_utilisation`, `gpu_temperature` and `gpu_memory_used`, and requires the public app id. A query that does not declare this selector rejects it rather than ignoring it. ### [appId](#get-metric-series-query-appid) - **Type**: `string[]` Restrict the list-scoped queries (`apps_request_volume`, `apps_error_volume`, `apps_request_duration`) to these app ids: one series per id, in request order, with all-null series for apps that had no samples. Values are quarter-hourly over `window=24h`. An id with no live app is dropped rather than refused: a deleted app has no traffic to pad. Repeat the parameter once per id on the current list page (at most 100, matching `listApps`). Omit it to receive every app in the organization that had data, unless that set is larger than this query will expand: then the call is a `422` on `appId` and the list page should name the apps it is showing. Other queries reject this parameter. ### [endpointId](#get-metric-series-query-endpointid) - **Type**: `string (uuid)[]` Restrict expand-by-`endpoint_id` queries (`endpoints_request_volume`, `endpoints_request_rate`, `endpoints_latency_p95`, `endpoints_latency_p99`, `endpoints_error_volume`, `endpoints_request_count`) to these endpoint ids: one series per id, in request order, with all-null series for endpoints that had no traffic. `endpoints_request_volume` is quarter-hourly request counts over `window=24h`. The others are latest-window scalars over `window=1h`. Buckets that started before that endpoint row's `createdAt` are null. Repeat the parameter once per id on the current list page (at most 100, matching `listEndpoints`). Omit it to receive every endpoint on the selected app that had data, unless that set is larger than this query will expand: then the call is a `422` on `endpointId` and the list page should name the endpoints it is showing. Other queries reject this parameter. Requires the public app id. ### [Response](#get-metric-series-response) ### [query](#get-metric-series-response-query) - **Type**: `string` - **Required**: true ### [unit](#get-metric-series-response-unit) - **Type**: `string` - **Required**: true ### [step\_s](#get-metric-series-response-step_s) - **Type**: `integer` - **Required**: true - **Format**: `int64` Seconds between adjacent timestamps in `t`. ### [aggregation](#get-metric-series-response-aggregation) - **Type**: `string` - **Required**: true ### [window](#get-metric-series-response-window) - **Type**: `object (2 properties)` - **Required**: true The range covered, as `(from, to]`. #### [from](#get-metric-series-response-window-from) - **Type**: `integer` - **Required**: true - **Format**: `int64` EXCLUSIVE: one `step_s` before the first timestamp in `t`. Nothing is sampled at `from` itself. #### [to](#get-metric-series-response-window-to) - **Type**: `integer` - **Required**: true - **Format**: `int64` INCLUSIVE, and equal to the last timestamp in `t`. ### [t](#get-metric-series-response-t) - **Type**: `integer[]` - **Required**: true The timestamp axis, shared by every series. Each value is the END of its bucket, in unix seconds: a point at `t[i]` is the aggregate over `(t[i] - step_s, t[i]]`, which is what the underlying range functions compute. ### [series](#get-metric-series-response-series) - **Type**: `object[] (3 properties)` - **Required**: true #### [id](#get-metric-series-response-series-id) - **Type**: `string` - **Required**: true Stable identifier for this line. For chart queries it does not change, so a client can map it to a color or legend position. For `apps_request_volume` it is the app id of that line. #### [label](#get-metric-series-response-series-label) - **Type**: `string` - **Required**: true Display text. May change without notice. Do not switch on it. #### [v](#get-metric-series-response-series-v) - **Type**: `number[]` - **Required**: true Values, positionally aligned to `t` and always the same length. `null` is a genuine absence of data, never a zero, and is sent explicitly rather than omitted so a gap draws as a gap instead of a line across an outage. ### [Errors](#get-metric-series-errors)
StatusWhen
400The request was malformed and could not be parsed (e.g. invalid JSON). A well-formed request that fails validation returns 422 instead.
401Missing or invalid credentials
404Resource not found
422The request was well-formed but semantically invalid (e.g. a missing or out-of-range field). A request that could not be parsed returns 400. The errors array carries one entry per offending field.
429The organization's read allowance is exhausted. Retry-After says how long to wait. A client that ignores it will keep being refused.
500Unexpected server error
502An upstream service was unreachable or refused the request
503A required service is temporarily unavailable
504An upstream service did not answer in time
## Read log entries GETapi.serverless.runware.ai/v1/logs/queries/{queryId}/entries Returns one page of log entries in the requested `sort` order, newest first by default, with opaque cursors for the neighboring pages when they exist. `nextCursor` continues in the `sort` order and `prevCursor` goes back against it, so a client can walk a window in either direction from either end. A cursor is only valid for the `sort` it was issued under. Reusing one under the other ordering returns `400`. Query ids and their supported selectors are listed by the insights catalog. > [!NOTE] > `nextCursor` continues in the `sort` order and `prevCursor` goes back against it, so you can walk a window from either end. A cursor is only valid for the `sort` it was issued under. ### [Request](#get-log-entries-request) Path ### [queryId](#get-log-entries-path-queryid) - **Type**: `string` - **Required**: true A query id from the catalog. Deliberately not an enum: the downstream registry is the source of ids, and an enum here would be a second list to keep in step with it. Query ### [window](#get-log-entries-query-window) - **Type**: `string` - **Required**: true The time window. A closed set rather than a free-form range, because every distinct range defeats the server-side cache alignment that makes a sliding window cheap. Only the windows a query lists in the catalog can be asked of it. **Allowed values**: `1h` `6h` `24h` `7d` `30d` ### [limit](#get-log-entries-query-limit) - **Type**: `integer` - **Format**: `int32` - **Min**: `1` - **Max**: `100` - **Default**: `20` Maximum number of items to return. ### [cursor](#get-log-entries-query-cursor) - **Type**: `string` Opaque pagination cursor returned by a previous call, as `nextCursor` or, on the operations that offer one, `prevCursor`. ### [sort](#get-log-entries-query-sort) - **Type**: `string` Ordering of a log page and the direction `nextCursor` moves in. Only `getLogEntries` accepts it. **Allowed values**: `newest` `oldest` ### [deployment](#get-log-entries-query-deployment) - **Type**: `string` - **Min**: `6` - **Max**: `30` Narrow to one app. ### [endpoint](#get-log-entries-query-endpoint) - **Type**: `string (uuid)` - **Format**: `UUID v4` Narrow to one endpoint within an app. The value is the allocated `endpoints.id` (UUID), never the customer-authored path. Accepted by `throughput`, `latency` and `error_rate`, and requires the public app id. A query that does not declare this selector rejects it rather than ignoring it. ### [Response](#get-log-entries-response) ### [entries](#get-log-entries-response-entries) - **Type**: `object[] (4 properties)` - **Required**: true #### [time](#get-log-entries-response-entries-time) - **Type**: `integer` - **Required**: true - **Format**: `int64` #### [level](#get-log-entries-response-entries-level) - **Type**: `string` #### [body](#get-log-entries-response-entries-body) - **Type**: `string` - **Required**: true #### [fields](#get-log-entries-response-entries-fields) - **Type**: `object` ### [nextCursor](#get-log-entries-response-nextcursor) - **Type**: `string` Opaque. Continues in the `sort` order. Absent on the last page. ### [prevCursor](#get-log-entries-response-prevcursor) - **Type**: `string` Opaque. Goes back against the `sort` order to the page before this one. Absent on the first page. Inside a group of entries sharing one timestamp, which has no order of its own, the pages read backwards can split the group differently from the pages read forwards. Each entry is still returned once either way. ### [Errors](#get-log-entries-errors)
StatusWhen
400The request was malformed and could not be parsed (e.g. invalid JSON). A well-formed request that fails validation returns 422 instead.
401Missing or invalid credentials
404Resource not found
422The request was well-formed but semantically invalid (e.g. a missing or out-of-range field). A request that could not be parsed returns 400. The errors array carries one entry per offending field.
429The organization's read allowance is exhausted. Retry-After says how long to wait. A client that ignores it will keep being refused.
500Unexpected server error
502An upstream service was unreachable or refused the request
503A required service is temporarily unavailable
504An upstream service did not answer in time
## Follow logs GETapi.serverless.runware.ai/v1/logs/queries/{queryId}/tail Streams new application log entries as Server-Sent Events. The stream sends keepalive comments while quiet and ends with an `end` event when its connection lifetime expires or the service shuts down. Clients should reconnect after an `end` event. Use the `runtime_tail` query id. Queries that sort or aggregate are rejected because they cannot be followed live. > [!NOTE] > Server-Sent Events, with keepalive comments while the stream is quiet. It ends with an `end` event when its connection lifetime expires, and the client is expected to reconnect. A live stream has no window, and a query that sorts or aggregates is rejected because it cannot be followed. ### [Request](#tail-log-entries-request) Path ### [queryId](#tail-log-entries-path-queryid) - **Type**: `string` - **Required**: true A query id from the catalog. Deliberately not an enum: the downstream registry is the source of ids, and an enum here would be a second list to keep in step with it. Query ### [deployment](#tail-log-entries-query-deployment) - **Type**: `string` - **Required**: true - **Min**: `6` - **Max**: `30` App whose new log entries are streamed. ### [Errors](#tail-log-entries-errors)
StatusWhen
400The request was malformed and could not be parsed (e.g. invalid JSON). A well-formed request that fails validation returns 422 instead.
401Missing or invalid credentials
404Resource not found
422The request was well-formed but semantically invalid (e.g. a missing or out-of-range field). A request that could not be parsed returns 400. The errors array carries one entry per offending field.
429The organization's read allowance is exhausted. Retry-After says how long to wait. A client that ignores it will keep being refused.
500Unexpected server error
502An upstream service was unreachable or refused the request
503A required service is temporarily unavailable
504An upstream service did not answer in time