Hi All,
To summarize, SESSION options are part of the query plan and distributed along with the query (not through ZK.) So, scenario 1 will always be fine. Since, for SESSION options there is only one distribution path, everything Just Works. Session options are set per connection, and tend to be few in number, so sending them with the query is simple and light-weight.
SESSION options are not part of the query plan and are instead distributed via ZK. If a SYSTEM option is set when no queries are running, things also Just Work. Prior to recent work, all options had values at the session level, and presumably we wanted to avoid the cost of serializing them across the network when we could just get them from ZK.
The comment about the race condition was simply an outcome of asking, what happens if we change a SYSTEM option in one connection concurrently with running a query in another? Since SYSTEM options spread throughout the cluster via the ZK path; and query fragments spread via the Foreman-to-Drillbit link, then they are not synchronized and can arrive a different times, in different orders, on different nodes.
However, after Jyothsna's changes last summer, we could capture just the set system options as well as just the set session options. (Look at FallbackOptionManager.getOptionList().) (This is based on a quick review of the code; perhaps I'm missing something. Easy to check: just look at the serialized query plan in the Drill log.)
So, yes, scenario 2 is a potential race condition, depending on when the option is checked. If the option is checked early in the query, the race condition window is small. If an option is checked later in the query (after running for 10 minutes to read a large table, say), then the window is large. Would have to analyze each use to see when each bit of code checks the option, but the race condition is inherent in the fact that we don't copy system options.
The race condition is a distraction however. The real point is Joe need not worry about problems with setting session options as they travel with the query fragments.
Thanks,
- Paul
On Wednesday, September 5, 2018, 1:44:29 PM PDT, Kunal Khatua <***@apache.org> wrote:
John
I think Paul's point is that during the planning & setup phase, the race condition can occur.Â
So, for scenario 1, as long as you have a SESSION var set to CSV, even if Paul comes along and changes the SYSTEM from parquet to json, you will be unaffected as your operator has the overriding var value already provided by the planner (guaranteed not to change).Â
For scenario 2, since you dont specify a SESSION var the operators can potentially see a race condition because the SYSTEM value can change.
So, yes, that is a lot of ZK reads. However, I'm assuming that Drillbits are caching the SYSTEM values and simply lsitening for changes (and reading them when notified).Â
~ Kunal
On 9/5/2018 12:01:00 PM, John Omernik <***@omernik.com> wrote:
Paul, let's talk about this race condition you mention
Let's use a real option here for clarity. store.format.
SYSTEM store.format is parquet
Scenario 1: I log on, I set SESSION store.format to csv and run CREATE
TABLE foo as select * from bar. The SESSION variable is read from my
login and that is passed to all drill bits that may get a fragment....
correct? I won't have some fragments on other drill bits reading SYSTEM of
parquet because I set a SESSION of csv.
Scenario 2 (Your described race condition): I logon in, I do NOT set a
SESSION variable for store.format. (Thus it's parquet). and run CREATE
TABLE foo as select * from bar. As it is running, and before the final
output, I change SYSTEM store.format to csv. And now I will run into this
race condition?
Scenario 2 is odd to me. Using the principle of least surprise, I would
expect to have my variables read once during planning and projected to all
downstream fragments rather than have a situation where variables could be
read at different times. This also seems like a lot of unnecessary ZK
reads... each fragment has to look up ZK settings? Should the foremen of
the query do that once? Are you sure this race condition exists? This
just doesn't sit right with me.
On Wed, Sep 5, 2018 at 11:35 AM, Paul Rogers
Post by Paul RogersHi Joe,
To answer our question about how options "travel"...
Drill maintains system options in ZK. Session options are maintained per
connection on the Foreman Drillbit to which the user connects. This is why
a simple round-robin load balancer does not work: why load balancing has to
be session-oriented as is the Drill JDBC an ODBC clients. (It is also why
the non-secure REST API does not preserve session options.) If you query
the session options table, you get a merger of the system options from ZK
overplayed with the session options on the Forman.
When a query executes, the query planner copies session options into the
query play that it sends to Drillbits for execution. The operators in each
Drillbit merge the session options (from the query plan) and the system
options (from ZK) to get the full set of options for the query.
Note that this does, in fact, create a race condition: if we change a
system option while queries run, some fragments of the query may see the
old value, others the new value. The workaround is simple: don't change
system options while queries are running.
Thanks,
- Paul
On Wednesday, September 5, 2018, 8:42:02 AM PDT, Joe Auty
Hmmm....
So, if user 1 sets the store.format to CSV on Drillbit 1 and work gets
farmed out to Drillbit 2, this session setting will "travel" with the
user from drillbit to drillbit? We were originally thinking that this
would be the case if the session information was retained in Zookeeper,
but we weren't so sure about this. How is session information tracked
and maintained across multiple drillbits?
Thanks so much for taking the time to engage on this!
Post by John OmernikRereading your post, I think there is some concern between
embedded/single
Post by John Omernikdrillbit mode, and distributed mode.
When you run multiple drillbits in distributed mode, you will (should) be
enabling authentication. Thus each user will log in to "a" drill bit.
There is no concern on which one, it's not the only one doing the work,
You
Post by John Omernikcould have two users on the same drill bit, or on different bits. The
"System" options will be set by default for all users on logon and at
this
Post by John Omernikpoint, a session for the user is created. You can see what the current
options for that user is with "Select * from sys.options" If you are a
user who is in the administrative group (security.admin.users in select *
from sys.options) then you can use
ALTER SYSTEM set `somekey` = `someval`
And that will set the default option for all users when they logon.
A user can use
ALTER SESSION set `somekey = `someval`
To alter what say their store.format is. So User 1 can authentication
and
Post by John Omernikconnect to Drillbit 1 or 2, and set store.format (at the session level)
to
Post by John Omernikparquet and user 2 can logon it Drillbit 1 or 2 (doesn't matter, even if
user 1 is logged into the same drill bit with a different store.format)
and
Post by John Omernikset it to csv.
I think the key here is authenticated distributed mode in Drill and that
will be how you will do what you need to do.
John
Post by John OmernikThe session is the users session, not the drill bit. Since you are
connected to a specific drill bit, when you alter session it will work.
Try
Post by John OmernikPost by John Omernikto use session stickiness or pinning on your HA solution and you will be
go. With my DNS round robin it picks a "connecting" drill bit and
sticks to
Post by John OmernikPost by John Omernikit until the session is done. The settings apply to the drill cluster
while
Post by John OmernikPost by John Omernikin distributed mode.
On Tue, Sep 4, 2018 at 3:56 PM, Joe Auty
Post by Joe AutyThanks for your response John!
We are using Drill both in an ETL context, as well as for general
warehouse queries. One Drill user uses store format set to Parquet
while
Post by John OmernikPost by John OmernikPost by Joe Autythe other uses store format set to CSV to read and write from HDFS. We
are
Post by John OmernikPost by John OmernikPost by Joe Autycurrently using Kubernetes Services rather than DNS round robin, but we
only have one drillbit in the cluster while we try to sort out this
issue.
Post by John OmernikPost by John OmernikPost by Joe AutyI'm not clear though how alter session would work with a session with
multiple drillbits involved? We need to ensure that the right store
format
Post by John OmernikPost by John OmernikPost by Joe Autyoption is set, so like I said we are using different Drill usernames
and
Post by John OmernikPost by John OmernikPost by Joe Autysessions to accommodate this, but how would alter session commands
apply to
Post by John OmernikPost by John OmernikPost by Joe Autypreserve these different settings across multiple drillbits?
Are these ETL ish type queries? store.format should only apply when
Drill
is writing data, when it is reading, it uses the filenames and other hints
to read.
Thus, if you do HA, say with DNS (like like in the other thread) and prior
to running your CREATE TABLE AS (I Am assuming this is what you are doing)
you can do ALTER SESSION set store.format = 'parquet'
Instead of setting the ALTER SYSTEM, you can use ALTER SESSION so it
only
Post by John OmernikPost by John OmernikPost by Joe Autyapplies to the current session, regardless of foreman.
John
On Tue, Sep 4, 2018 at 1:00 PM, Joe Auty
Hello,
Post by Joe AutyWe need to have some queries executed with store.format set to
parquet
Post by John OmernikPost by John OmernikPost by Joe AutyPost by Joe Autyand
some with this option set to CSV. To date we have experimented with setting
the store format for sessions controlled by using two separate user logins
as a sort of context switch, but I'm wondering if the group here
might
will
Post by John OmernikPost by John OmernikPost by Joe AutyPost by Joe Autyscale a little better for us?
The main problem we have with this approach is in introducing
multiple
are
session
(which
Post by John OmernikPost by John OmernikPost by Joe AutyPost by Joe AutyI've
chosen not to hijack) about HA Drill from a more general standpoint,
you
for
options.
Post by John OmernikPost by John OmernikPost by Joe AutyPost by Joe AutyHere are the various attempts and approaches we have come up with so far.
I'm wondering if you'd have any general advice as to which approach would
be best for us to take, considering future plans for Drill itself.
For
Post by John OmernikPost by John OmernikPost by Joe AutyPost by Joe Autyexample, if need be we can write our own plugin(s) if this is the smartest
- embedded the store.format option into the query itself by chaining
multiple queries together separated by a comma (it appears that this
doesn't work at all)
- look into writing some sort of plugin to allow us to scale our
current
Post by John OmernikPost by John OmernikPost by Joe AutyPost by Joe Autyapproach somehow (I realize that this is vague)
- a "foreman" approach where we stick with our current approach and direct
all requests to our "foreman"/master with the hope and expectation
that
Post by John OmernikPost by John OmernikPost by Joe AutyPost by Joe Autyit
will farm out work to the workers/slaves
- multiple clusters set with different settings
Each of these approaches seems to have its pros and cons. To
what approach do you think would be the smartest and most
future-proof