This reduces memory consumption for new idle subscriptions significantly
(see linked ticket).
The hypothesis is: we fork a lot of threads per websocket, and some of
these use slightly more than the initial 1K stack size, so the first
overflow balloons to 32K, when significantly less is required.
However: running with `+RTS -K1K -xc` did not seem to show evidence of
any overflows! So it's a mystery why this improves things.
GHC should probably also be doubling the stack buffer at each overflow
or doing something even smarter; the knobs we have aren't so helpful.
Introspection query is failing with `type info not found for xxxx` error message if multiple actions are defined with reused PG scalars. The fix for the same.
* Benchmark GraphQL queries using wrk
* fix console assets dir
* Store wrk parameters as well
* Add details about storing results in Readme
* Remove files in bench-wrk while computing server shasum
* Instead of just getting maximum throughput per query per version,
create plots using wrk2 for a given set of requests per second.
The maximum throughput is used to see what values of requests per second are feasible.
* Add id for version dropdown
* Allow specifiying env and args for GraphQL Engine
1) Arguments defined after -- will be applied as arguments to Hasura GraphQL Engine
2) Script will also pass the environmental variables to Hasura GraphQL Engine instances
Hasura GraphQL engine can be run with the given environmental variables and arguments as follows
$ export HASURA_GRAPHQL_...=....
$ python3 hge_wrk_bench.py -- --hge_arg1 val1 --hge_arg2 val2 ...
* Use matplotlib instead of plotly for figures
* Show throughput graph also.
It maybe useful in checking performance regression across versions
* Support storing results in s3
Use --upload-root-uri 's3://bucket/path' to upload results inside the
given path.When specified, the results will be uploaded to the bucket,
including latencies, latency histogram, and the test setup info.
The s3 credentials should be provided as given in AWS boto3 documentation.
* Allow specifying a name for the test scenario
* Fix open latency uri bug
* Update wrk docker image
* Keep ylim a little higher than maximum so that the throughput plot is clearly visible
* Show throughput plots for multiple queries at the same time
* 1) Adjust size of dropdowns
2) Make label for requests/sec invisible when plot type is throughput
* 1) Adding boto3 to requirements.txt
2) Removing CPU Key print line
3) Adding info about the tests that will be run with wrk2
* Docker builder fo wrk-websocket-server
* Make it optional to setup remote graphql-engine
* Listen on all interfaces and enable ping thread
* Add bench_scripts to wrk-websocket-server docker
* Use 127.0.0.1 instead of 'localhost' to address local hge
For some reason it seems wrk was hanging trying to resolve 'localhost'.
ping was able to fine from the same container, so I'm not sure what the
deal was. Probably some local misconfiguration on my machine, but maybe
this change will also help others.
* Store latency samples in subdirectory, server_shasum just once at start, additional docs
* Add a note on running the benchmarks in the simplest way
* Add a new section on how to run benchmarks on a new linux hosted instance
Co-authored-by: Nizar Malangadan <nizar-m@users.noreply.github.com>
Co-authored-by: Brandon Simmons <brandon.m.simmons@gmail.com>
Co-authored-by: Karthikeyan Chinnakonda <karthikeyan@hasura.io>
Co-authored-by: Brandon Simmons <brandon@hasura.io>
Co-authored-by: Vamshi Surabhi <0x777@users.noreply.github.com>
* new typeclass to abstract the logic of QueryLog-ing
* abstract the logic of logging websocket-server logs
introduce a MonadWSLog typeclass
* move catalog initialization to init step
expose a helper function to migrate catalog
create schema cache in initialiseCtx
* expose various modules and functions for pro
* generalize PGExecCtx to support specialized functions for various operations
* fix tests compilation
* allow customising PGExecCtx when starting the web server
* fix relay introspection failing if any views exist, fix#5020
* reduce base64 encoded node id length, close#5037
* make node field type non-nullable in an edge
* more relay tests with permissions & complete restructure of test yaml files
Co-authored-by: Aravind <aravindkp@outlook.in>
Co-authored-by: Vamshi Surabhi <0x777@users.noreply.github.com>
The bulk of changes here is some shifting of code around and a little
parameterizing of functions for easier testing.
Also: comments, some renaming for clarity/less-chance-for-misue.
Store the admin secret only as a hash to prevent leaking the secret
inadvertently, and to prevent timing attacks on the secret.
NOTE: best practice for stored user passwords is a function with a
tunable cost like bcrypt, but our threat model is quite different (even
if we thought we could reasonably protect the secret from an attacker
who could read arbitrary regions of memory), and bcrypt is far too slow
(by design) to perform on each request. We'd have to rely on our
(technically savvy) users to choose high entropy passwords in any case.
Referencing #4736
* validation support for unions and interfaces
* refactor SQL generation logic for improved readability
* '/v1/relay' endpoint for relay schema
* implement 'Node' interface and top level 'node' field resolver
* add relay toggle on graphiql
* fix explain api response & index plan id with query type
* add hasura mutations to relay
* add relay pytests
* update CHANGELOG.md
Co-authored-by: rakeshkky <12475069+rakeshkky@users.noreply.github.com>
Co-authored-by: Rishichandra Wawhal <rishi@hasura.io>
Co-authored-by: Rikin Kachhia <54616969+rikinsk@users.noreply.github.com>
* changes for poller-log
add various multiplexed query info in poller-log
* minor cleanup, also fixes a bug which will return duplicate data
* Live query poller stats can now be logged
This also removes in-memory stats that are collected about batched
query execution as the log lines when piped into an monitoring tool
will give us better insights.
* allow poller-log to be configurable
* log minimal information in the livequery-poller-log
Other information can be retrieved from /dev/subscriptions/extended
* fix few review comments
* avoid marshalling and unmarshalling from ByteString to EncJSON
* separate out SubscriberId and SubscriberMetadata
Co-authored-by: Anon Ray <rayanon004@gmail.com>
When compiling the graphql-engine binary with `-O2`, ghc-8.10 seems to
be stuck at the module `Server.Init` while consuming `17G` of RAM (for 5
minutes at least before I forcefully terminated the compilation). With
this pragma, ghc-8.10 now takes under `12G` to compile graphql-engine
binary.
Introspection queries accept variables, but we need to make sure to
also touch the variables that we ignore, so that an introspection
query is marked not reusable if we are not able to build a correct
query plan for it.
A better solution here would be to deal with such unused variables
correctly, so that more introspection queries become reusable.
An even better solution would be to type-safely track *how* to reuse
which variables, rather than to split the reusage marking from the
planning.
Co-authored-by: Tirumarai Selvan <tiru@hasura.io>
This also seems to squash a stubborn space leak we see with
subscriptions (linking to canonical #3388 for reference).
This may also fix some of the "Unexpected exception" websockets
exceptions we are now surfacing (see e.g. #4344)
Also: dev.sh: fix hpc reporting
Initial work on this done by Vamshi.
There are two implementations of a Cache, namely a bounded and an
unbounded variant. This can be elegantly captured in a type class.
In addition to reducing the amount of error-prone code in the
definition of the cache, this version reduces the amount of
error-prone code in usage sites of the cache, as it makes the cache
into an abstract object, so that a calling site cannot distinguish
between cache types. Any decision about what should be cached should
be made through the interface of a cache, rather than at the callsite,
and this is captured by this variant.
* Add support for multiple top-level fields in a subscription to improve testability of subscriptions
* Add an internal flag to enable multiple subscriptions
* Add missing call to withConstructorFn in live queries (fix#3239)
Co-authored-by: Alexis King <lexi.lambda@gmail.com>
