The previous name, 'johncolvin/bluetradingsystems', was chosen in a
misguided effort to adhere to a naming convention recommended by Docker
Hub (https://docs.docker.com/docker-hub/repos/create/). In hindsight,
this convention isn't well suited to our use case. Our images are shared
by multiple users (e.g., sparkd dev users all use the same sparkd-dev
image), so it doesn't make sense to name the repository after an
arbitrary user.
MarketDataGateway manages emumba/deribit connections now
Bugs (fixed):
* An spdlog exception is thrown because the run_emumba thread spawns a
2nd emumba_logger.
Cause: The first emumba logger is spawned by the EmMDListener
instantiated in MarketDataGateway. We shouldn't be starting emumba
connections in 2 places.
Leave the image's default user as 'root' so we can modify the non-root user's
uid/gid. This is a common technique used in containers to gain file permissions
in mounted dirs.
The default user's uid/gid can't be changed because a bash process runs under
their ownership as long as the container is up. A user's uid/gid cannot be
changed while that user owns a process (at least not via 'usermod').
Bugs (fixed):
* The default user does not have sudo permission in the container,
unless the user is specified explicitly in the 'docker exec' cmd
(via -u <user>).
Cause: When the default user is specified with a specific group
(in this case, its namesake 'bts'), it is excluded from all other
groups, including 'sudo'.
This is a more accurate and appropriate name than 'spark-dev' because
the image is designed to build/run Sparkd and the Spark Api Server (both
of which are daemons, hence the appropriateness of the 'd' suffix). It
isn't designed to build/run the front-end desktop application, Spark.
Bugs (fixed):
* Assert crash in MDSource.h "Marketdata heartbeats are bad..."
Cause: The MD cleanup procedure, which is supposed to delete md files
from shared memory (from previous runs), wasn't deleting anything.
The file globbing pattern it was using (to determine which files to
delete) never produced any matches because it had the wrong prefix. As
a result, market data would often be loaded with a stale heartbeat
timestamp from a previous run.
Note: This occurred as a side effect of making the shared memory prefix
configurable when we ported the api server over to Linux (so the api
server and sparkd could run on the same container/machine whilst
maintaining independent shared memory files).