dendrite/docs/peeking.md

26 lines
2 KiB
Markdown
Raw Permalink Normal View History

## Peeking
Peeking over federation via MSC2444 (#1391) * a very very WIP first cut of peeking via MSC2753. doesn't yet compile or work. needs to actually add the peeking block into the sync response. checking in now before it gets any bigger, and to gather any initial feedback on the vague shape of it. * make PeekingDeviceSet private * add server_name param * blind stab at adding a `peek` section to /sync * make it build * make it launch * add peeking to getResponseWithPDUsForCompleteSync * cancel any peeks when we join a room * spell out how to runoutside of docker if you want speed * fix SQL * remove unnecessary txn for SelectPeeks * fix s/join/peek/ cargocult fail * HACK: Track goroutine IDs to determine when we write by the wrong thread To use: set `DENDRITE_TRACE_SQL=1` then grep for `unsafe` * Track partition offsets and only log unsafe for non-selects * Put redactions in the writer goroutine * Update filters on writer goroutine * wrap peek storage in goid hack * use exclusive writer, and MarkPeeksAsOld more efficiently * don't log ascii in binary at sql trace... * strip out empty roomd deltas * re-add txn to SelectPeeks * re-add accidentally deleted field * reject peeks for non-worldreadable rooms * move perform_peek * fix package * correctly refactor perform_peek * WIP of implementing MSC2444 * typo * Revert "Merge branch 'kegan/HACK-goid-sqlite-db-is-locked' into matthew/peeking" This reverts commit 3cebd8dbfbccdf82b7930b7b6eda92095ca6ef41, reversing changes made to ed4b3a58a7855acc43530693cc855b439edf9c7c. * (almost) make it build * clean up bad merge * support SendEventWithState with optional event * fix build & lint * fix build & lint * reinstate federated peeks in the roomserver (doh) * fix sql thinko * todo for authenticating state returned by /peek * support returning current state from QueryStateAndAuthChain * handle SS /peek * reimplement SS /peek to prod the RS to tell the FS about the peek * rename RemotePeeks as OutboundPeeks * rename remote_peeks_table as outbound_peeks_table * add perform_handle_remote_peek.go * flesh out federation doc * add inbound peeks table and hook it up * rename ambiguous RemotePeek as InboundPeek * rename FSAPI's PerformPeek as PerformOutboundPeek * setup inbound peeks db correctly * fix api.SendEventWithState with no event * track latestevent on /peek * go fmt * document the peek send stream race better * fix SendEventWithRewrite not to bail if handed a non-state event * add fixme * switch SS /peek to use SendEventWithRewrite * fix comment * use reverse topo ordering to find latest extrem * support postgres for federated peeking * go fmt * back out bogus go.mod change * Fix performOutboundPeekUsingServer * Fix getAuthChain -> GetAuthChain * Fix build issues * Fix build again * Fix getAuthChain -> GetAuthChain * Don't repeat outbound peeks for the same room ID to the same servers * Fix lint * Don't omitempty to appease sytest Co-authored-by: Kegan Dougal <kegan@matrix.org> Co-authored-by: Neil Alexander <neilalexander@users.noreply.github.com>
2021-01-22 14:55:08 +00:00
Local peeking is implemented as per [MSC2753](https://github.com/matrix-org/matrix-doc/pull/2753).
Implementationwise, this means:
* Users call `/peek` and `/unpeek` on the clientapi from a given device.
* The clientapi delegates these via HTTP to the roomserver, which coordinates peeking in general for a given room
* The roomserver writes an NewPeek event into the kafka log headed to the syncserver
Peeking over federation via MSC2444 (#1391) * a very very WIP first cut of peeking via MSC2753. doesn't yet compile or work. needs to actually add the peeking block into the sync response. checking in now before it gets any bigger, and to gather any initial feedback on the vague shape of it. * make PeekingDeviceSet private * add server_name param * blind stab at adding a `peek` section to /sync * make it build * make it launch * add peeking to getResponseWithPDUsForCompleteSync * cancel any peeks when we join a room * spell out how to runoutside of docker if you want speed * fix SQL * remove unnecessary txn for SelectPeeks * fix s/join/peek/ cargocult fail * HACK: Track goroutine IDs to determine when we write by the wrong thread To use: set `DENDRITE_TRACE_SQL=1` then grep for `unsafe` * Track partition offsets and only log unsafe for non-selects * Put redactions in the writer goroutine * Update filters on writer goroutine * wrap peek storage in goid hack * use exclusive writer, and MarkPeeksAsOld more efficiently * don't log ascii in binary at sql trace... * strip out empty roomd deltas * re-add txn to SelectPeeks * re-add accidentally deleted field * reject peeks for non-worldreadable rooms * move perform_peek * fix package * correctly refactor perform_peek * WIP of implementing MSC2444 * typo * Revert "Merge branch 'kegan/HACK-goid-sqlite-db-is-locked' into matthew/peeking" This reverts commit 3cebd8dbfbccdf82b7930b7b6eda92095ca6ef41, reversing changes made to ed4b3a58a7855acc43530693cc855b439edf9c7c. * (almost) make it build * clean up bad merge * support SendEventWithState with optional event * fix build & lint * fix build & lint * reinstate federated peeks in the roomserver (doh) * fix sql thinko * todo for authenticating state returned by /peek * support returning current state from QueryStateAndAuthChain * handle SS /peek * reimplement SS /peek to prod the RS to tell the FS about the peek * rename RemotePeeks as OutboundPeeks * rename remote_peeks_table as outbound_peeks_table * add perform_handle_remote_peek.go * flesh out federation doc * add inbound peeks table and hook it up * rename ambiguous RemotePeek as InboundPeek * rename FSAPI's PerformPeek as PerformOutboundPeek * setup inbound peeks db correctly * fix api.SendEventWithState with no event * track latestevent on /peek * go fmt * document the peek send stream race better * fix SendEventWithRewrite not to bail if handed a non-state event * add fixme * switch SS /peek to use SendEventWithRewrite * fix comment * use reverse topo ordering to find latest extrem * support postgres for federated peeking * go fmt * back out bogus go.mod change * Fix performOutboundPeekUsingServer * Fix getAuthChain -> GetAuthChain * Fix build issues * Fix build again * Fix getAuthChain -> GetAuthChain * Don't repeat outbound peeks for the same room ID to the same servers * Fix lint * Don't omitempty to appease sytest Co-authored-by: Kegan Dougal <kegan@matrix.org> Co-authored-by: Neil Alexander <neilalexander@users.noreply.github.com>
2021-01-22 14:55:08 +00:00
* The syncserver tracks the existence of the local peek in the syncapi_peeks table in its DB, and then starts waking up the peeking devices for the room in question, putting it in the `peek` section of the /sync response.
Peeking over federation via MSC2444 (#1391) * a very very WIP first cut of peeking via MSC2753. doesn't yet compile or work. needs to actually add the peeking block into the sync response. checking in now before it gets any bigger, and to gather any initial feedback on the vague shape of it. * make PeekingDeviceSet private * add server_name param * blind stab at adding a `peek` section to /sync * make it build * make it launch * add peeking to getResponseWithPDUsForCompleteSync * cancel any peeks when we join a room * spell out how to runoutside of docker if you want speed * fix SQL * remove unnecessary txn for SelectPeeks * fix s/join/peek/ cargocult fail * HACK: Track goroutine IDs to determine when we write by the wrong thread To use: set `DENDRITE_TRACE_SQL=1` then grep for `unsafe` * Track partition offsets and only log unsafe for non-selects * Put redactions in the writer goroutine * Update filters on writer goroutine * wrap peek storage in goid hack * use exclusive writer, and MarkPeeksAsOld more efficiently * don't log ascii in binary at sql trace... * strip out empty roomd deltas * re-add txn to SelectPeeks * re-add accidentally deleted field * reject peeks for non-worldreadable rooms * move perform_peek * fix package * correctly refactor perform_peek * WIP of implementing MSC2444 * typo * Revert "Merge branch 'kegan/HACK-goid-sqlite-db-is-locked' into matthew/peeking" This reverts commit 3cebd8dbfbccdf82b7930b7b6eda92095ca6ef41, reversing changes made to ed4b3a58a7855acc43530693cc855b439edf9c7c. * (almost) make it build * clean up bad merge * support SendEventWithState with optional event * fix build & lint * fix build & lint * reinstate federated peeks in the roomserver (doh) * fix sql thinko * todo for authenticating state returned by /peek * support returning current state from QueryStateAndAuthChain * handle SS /peek * reimplement SS /peek to prod the RS to tell the FS about the peek * rename RemotePeeks as OutboundPeeks * rename remote_peeks_table as outbound_peeks_table * add perform_handle_remote_peek.go * flesh out federation doc * add inbound peeks table and hook it up * rename ambiguous RemotePeek as InboundPeek * rename FSAPI's PerformPeek as PerformOutboundPeek * setup inbound peeks db correctly * fix api.SendEventWithState with no event * track latestevent on /peek * go fmt * document the peek send stream race better * fix SendEventWithRewrite not to bail if handed a non-state event * add fixme * switch SS /peek to use SendEventWithRewrite * fix comment * use reverse topo ordering to find latest extrem * support postgres for federated peeking * go fmt * back out bogus go.mod change * Fix performOutboundPeekUsingServer * Fix getAuthChain -> GetAuthChain * Fix build issues * Fix build again * Fix getAuthChain -> GetAuthChain * Don't repeat outbound peeks for the same room ID to the same servers * Fix lint * Don't omitempty to appease sytest Co-authored-by: Kegan Dougal <kegan@matrix.org> Co-authored-by: Neil Alexander <neilalexander@users.noreply.github.com>
2021-01-22 14:55:08 +00:00
Peeking over federation is implemented as per [MSC2444](https://github.com/matrix-org/matrix-doc/pull/2444).
Peeking over federation via MSC2444 (#1391) * a very very WIP first cut of peeking via MSC2753. doesn't yet compile or work. needs to actually add the peeking block into the sync response. checking in now before it gets any bigger, and to gather any initial feedback on the vague shape of it. * make PeekingDeviceSet private * add server_name param * blind stab at adding a `peek` section to /sync * make it build * make it launch * add peeking to getResponseWithPDUsForCompleteSync * cancel any peeks when we join a room * spell out how to runoutside of docker if you want speed * fix SQL * remove unnecessary txn for SelectPeeks * fix s/join/peek/ cargocult fail * HACK: Track goroutine IDs to determine when we write by the wrong thread To use: set `DENDRITE_TRACE_SQL=1` then grep for `unsafe` * Track partition offsets and only log unsafe for non-selects * Put redactions in the writer goroutine * Update filters on writer goroutine * wrap peek storage in goid hack * use exclusive writer, and MarkPeeksAsOld more efficiently * don't log ascii in binary at sql trace... * strip out empty roomd deltas * re-add txn to SelectPeeks * re-add accidentally deleted field * reject peeks for non-worldreadable rooms * move perform_peek * fix package * correctly refactor perform_peek * WIP of implementing MSC2444 * typo * Revert "Merge branch 'kegan/HACK-goid-sqlite-db-is-locked' into matthew/peeking" This reverts commit 3cebd8dbfbccdf82b7930b7b6eda92095ca6ef41, reversing changes made to ed4b3a58a7855acc43530693cc855b439edf9c7c. * (almost) make it build * clean up bad merge * support SendEventWithState with optional event * fix build & lint * fix build & lint * reinstate federated peeks in the roomserver (doh) * fix sql thinko * todo for authenticating state returned by /peek * support returning current state from QueryStateAndAuthChain * handle SS /peek * reimplement SS /peek to prod the RS to tell the FS about the peek * rename RemotePeeks as OutboundPeeks * rename remote_peeks_table as outbound_peeks_table * add perform_handle_remote_peek.go * flesh out federation doc * add inbound peeks table and hook it up * rename ambiguous RemotePeek as InboundPeek * rename FSAPI's PerformPeek as PerformOutboundPeek * setup inbound peeks db correctly * fix api.SendEventWithState with no event * track latestevent on /peek * go fmt * document the peek send stream race better * fix SendEventWithRewrite not to bail if handed a non-state event * add fixme * switch SS /peek to use SendEventWithRewrite * fix comment * use reverse topo ordering to find latest extrem * support postgres for federated peeking * go fmt * back out bogus go.mod change * Fix performOutboundPeekUsingServer * Fix getAuthChain -> GetAuthChain * Fix build issues * Fix build again * Fix getAuthChain -> GetAuthChain * Don't repeat outbound peeks for the same room ID to the same servers * Fix lint * Don't omitempty to appease sytest Co-authored-by: Kegan Dougal <kegan@matrix.org> Co-authored-by: Neil Alexander <neilalexander@users.noreply.github.com>
2021-01-22 14:55:08 +00:00
For requests to peek our rooms ("inbound peeks"):
* Remote servers call `/peek` on federationapi
* The federationapi queries the federationsender to check if this is renewing an inbound peek or not.
* If not, it hits the PerformInboundPeek on the roomserver to ask it for the current state of the room.
* The roomserver atomically (in theory) adds a NewInboundPeek to its kafka stream to tell the federationserver to start peeking.
* The federationsender receives the event, tracks the inbound peek in the federationsender_inbound_peeks table, and starts sending events to the peeking server.
* The federationsender evicts stale inbound peeks which haven't been renewed.
For peeking into other server's rooms ("outbound peeks"):
* The `roomserver` will kick the `federationsender` much as it does for a federated `/join` in order to trigger a federated outbound `/peek`
* The `federationsender` tracks the existence of the outbound peek in in its federationsender_outbound_peeks table.
* The `federationsender` regularly renews the remote peek as long as there are still peeking devices syncing for it.
* TBD: how do we tell if there are no devices currently syncing for a given peeked room? The syncserver needs to tell the roomserver
somehow who then needs to warn the federationsender.