Deploy several of the project's services together

A path, beginning with /: a GET of it on the app's port is answered with a status from 200 to 399.

The app's port takes a connection.

A command that succeeds in the instance. It is run by the instance's shell: a list of arguments is written as one line for it.

An even number of MiB (a whole number of the 2 MiB pages a node loads a machine's memory in), in bytes. Any other size is 400 invalid_field (param size.memory_bytes): no node would make an instance of it, as none makes a sandbox of it (the node's SandboxLimitsCreate.memory_bytes).

There only where it was given as true.

How many instances of the app are kept awake in a region, whether a request comes or not: the floor. The nodes of the region that serve the app share it as they share max_instances (two nodes and a floor of 2: one on each; a floor of 1: one, on the node that is tried first; a floor of 5: three and two), and a node that is away is made up for by the others. A node makes its part when it is told, not at the first request, and neither stops nor deletes those instances for being idle; they are metered as instances that are awake. 0, which is what it is when nothing is said: none. Lowered, the instances beyond it are idle from that moment, and are stopped and deleted as idle instances are. At most max_instances, where that is given, and at most what the account's plan lets an app ask for in a region (caps.app_instances.limit: 5, 50 and 500): more is refused, 403 plan_cap_reached.

How many instances the app has at the most in a region: the nodes of the region that serve the app share it. At most what the account's plan lets an app ask for (caps.app_instances.limit: 5 on Free, 500 on Paid): more is refused, 403 plan_cap_reached, and what an app asked for on a plan that let it is kept, and held to the plan's number. null: the plan's number for an app that names none (caps.app_instances.default: 2 on Free, 50 on Paid), or min_instances where that is more. What the nodes are told is no more than the account has room for awake beside everything else of it that is awake, nor less than the app has awake.

A cap on what one instance has in hand at a time: the requests the edges sent it, and the connections it has open on its project's network (through its service's address, which the node shares out among the instances connection by connection, or to its own). An instance with that many is full, and another is made for more. null, which is what it is when nothing is said: no cap. A node then adds an instance when the ones it has are near their size in CPU or memory, and an app does not have to guess a number; a service that takes one request at a time gives one.

The most an instance waits without a request before it stops, in milliseconds. Below it, its node picks the wait by the app's own traffic: the wait that would have cost least over the app's last gaps between a request's end and the next one, a second up and idle against a stop and the resume after it in CPU time, so that an app whose requests come closer together than a stop pays for stays up through them, and one whose requests are far apart stops as soon as its last answer has ended (0: stops at once, always). null, which is what it is when nothing is said: the platform's most, three seconds. An app that said 100, the default before 2026-10-28, says null from then on. One that stays stopped is deleted by its node after delete_after_ms, and the next request that needs an instance gets a new one; the instances min_instances keeps awake are neither stopped nor deleted.

How long before an instance stops for being idle the programs connected to its guest's /run/sbx/awake are sent the line sleep N, in milliseconds: they flush what they hold, close their connections, or raise a hold, which puts the stop off. The stop comes that long after, whether they answer or not; where nobody is connected, it is not put off. A keyed instance is told as well. 0 or null: not told (the default).

How long the holds a program in an instance raises keep it up after its time idle has run out, in milliseconds: a program connects to /run/sbx/awake and sends + before it answers a request whose work goes on after the answer, and - (or closes the connection, or dies) when that work is done; the instance stops sleep_after_ms after the last hold is lowered. Past this time it stops whatever holds it, the programs connected told sleep N cap first where sleep_notice_ms asks for a notice. Held, it runs and is metered as running. 0: holds are not heeded. Null: 30000 (the default).

After how long idle an instance is deleted, counted from where its stop is counted from: the next request that needs one gets a new instance, a clone of the version's template, without what the deleted one had gathered in its memory and on its disk (keep what is to last in a volume or a database). A second to an hour, the platform's time, which is the most: more is refused (400 invalid_field), and an instance that is to stay warm is one of min_instances. null, which is what it is when nothing is said: the hour of the nodes (--app-delete-after-seconds). A change holds for the instances the app has as for the ones made after: each node judges every instance by the app's time as it is, once a second at the least. Not for the app's keyed instances, which keys.delete_after_idle_ms is for.

How long a stopped instance stays paused before its memory is saved and it sleeps for real, in milliseconds: paused, the next request is answered in a millisecond or two and its memory is metered; asleep, it wakes in a few more and only what it keeps is. Its node holds it to the longest it holds any instance paused (ten minutes); the host's memory running short, a slot another machine needs and a sleep somebody asks for put it to sleep sooner. Null (the default): its node's longest. A database's suspend_after_seconds is its app's.

A time of day in UTC, HH:MM, at which the app's keyed instances and its stateful service's instance that rest are woken every day, for work of their own that waits for a wake (a database's maintenance, say); the instances the nodes make and delete by the app's requests are woken by requests alone. The wake is no activity: what runs then is, and once that is done the instance rests again. A stateful service's instance is woken besides for the next run of its PostgreSQL server's pg_cron jobs, whatever this says. Null (the default): none. A database's wake_daily_at is its app's.

Ports no connection reaches, TCP or UDP, wherever it goes and whatever else the rules say. A Browsers session's always hold 25, 465, 587 and the ports mining pools listen on (3333, 4444, 5555, 7777, 14444, 45700): the control plane adds them to whatever a change of its rules gives.

There only where it was given as true.

A path, beginning with /: a GET of it on the app's port is answered with a status from 200 to 399.

The app's port takes a connection.

A command that succeeds in the instance. It is run by the instance's shell: a list of arguments is written as one line for it.

An even number of MiB (a whole number of the 2 MiB pages a node loads a machine's memory in), in bytes. Any other size is 400 invalid_field (param size.memory_bytes): no node would make an instance of it, as none makes a sandbox of it (the node's SandboxLimitsCreate.memory_bytes).

There only where it was given as true.

How many instances of the app are kept awake in a region, whether a request comes or not: the floor. The nodes of the region that serve the app share it as they share max_instances (two nodes and a floor of 2: one on each; a floor of 1: one, on the node that is tried first; a floor of 5: three and two), and a node that is away is made up for by the others. A node makes its part when it is told, not at the first request, and neither stops nor deletes those instances for being idle; they are metered as instances that are awake. 0, which is what it is when nothing is said: none. Lowered, the instances beyond it are idle from that moment, and are stopped and deleted as idle instances are. At most max_instances, where that is given, and at most what the account's plan lets an app ask for in a region (caps.app_instances.limit: 5, 50 and 500): more is refused, 403 plan_cap_reached.

How many instances the app has at the most in a region: the nodes of the region that serve the app share it. At most what the account's plan lets an app ask for (caps.app_instances.limit: 5 on Free, 500 on Paid): more is refused, 403 plan_cap_reached, and what an app asked for on a plan that let it is kept, and held to the plan's number. null: the plan's number for an app that names none (caps.app_instances.default: 2 on Free, 50 on Paid), or min_instances where that is more. What the nodes are told is no more than the account has room for awake beside everything else of it that is awake, nor less than the app has awake.

A cap on what one instance has in hand at a time: the requests the edges sent it, and the connections it has open on its project's network (through its service's address, which the node shares out among the instances connection by connection, or to its own). An instance with that many is full, and another is made for more. null, which is what it is when nothing is said: no cap. A node then adds an instance when the ones it has are near their size in CPU or memory, and an app does not have to guess a number; a service that takes one request at a time gives one.

The most an instance waits without a request before it stops, in milliseconds. Below it, its node picks the wait by the app's own traffic: the wait that would have cost least over the app's last gaps between a request's end and the next one, a second up and idle against a stop and the resume after it in CPU time, so that an app whose requests come closer together than a stop pays for stays up through them, and one whose requests are far apart stops as soon as its last answer has ended (0: stops at once, always). null, which is what it is when nothing is said: the platform's most, three seconds. An app that said 100, the default before 2026-10-28, says null from then on. One that stays stopped is deleted by its node after delete_after_ms, and the next request that needs an instance gets a new one; the instances min_instances keeps awake are neither stopped nor deleted.

How long before an instance stops for being idle the programs connected to its guest's /run/sbx/awake are sent the line sleep N, in milliseconds: they flush what they hold, close their connections, or raise a hold, which puts the stop off. The stop comes that long after, whether they answer or not; where nobody is connected, it is not put off. A keyed instance is told as well. 0 or null: not told (the default).

How long the holds a program in an instance raises keep it up after its time idle has run out, in milliseconds: a program connects to /run/sbx/awake and sends + before it answers a request whose work goes on after the answer, and - (or closes the connection, or dies) when that work is done; the instance stops sleep_after_ms after the last hold is lowered. Past this time it stops whatever holds it, the programs connected told sleep N cap first where sleep_notice_ms asks for a notice. Held, it runs and is metered as running. 0: holds are not heeded. Null: 30000 (the default).

After how long idle an instance is deleted, counted from where its stop is counted from: the next request that needs one gets a new instance, a clone of the version's template, without what the deleted one had gathered in its memory and on its disk (keep what is to last in a volume or a database). A second to an hour, the platform's time, which is the most: more is refused (400 invalid_field), and an instance that is to stay warm is one of min_instances. null, which is what it is when nothing is said: the hour of the nodes (--app-delete-after-seconds). A change holds for the instances the app has as for the ones made after: each node judges every instance by the app's time as it is, once a second at the least. Not for the app's keyed instances, which keys.delete_after_idle_ms is for.

How long a stopped instance stays paused before its memory is saved and it sleeps for real, in milliseconds: paused, the next request is answered in a millisecond or two and its memory is metered; asleep, it wakes in a few more and only what it keeps is. Its node holds it to the longest it holds any instance paused (ten minutes); the host's memory running short, a slot another machine needs and a sleep somebody asks for put it to sleep sooner. Null (the default): its node's longest. A database's suspend_after_seconds is its app's.

A time of day in UTC, HH:MM, at which the app's keyed instances and its stateful service's instance that rest are woken every day, for work of their own that waits for a wake (a database's maintenance, say); the instances the nodes make and delete by the app's requests are woken by requests alone. The wake is no activity: what runs then is, and once that is done the instance rests again. A stateful service's instance is woken besides for the next run of its PostgreSQL server's pg_cron jobs, whatever this says. Null (the default): none. A database's wake_daily_at is its app's.

Ports no connection reaches, TCP or UDP, wherever it goes and whatever else the rules say. A Browsers session's always hold 25, 465, 587 and the ports mining pools listen on (3333, 4444, 5555, 7777, 14444, 45700): the control plane adds them to whatever a change of its rules gives.

There only where it was given as true.