You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@calcite.apache.org by F21 <f2...@gmail.com> on 2017/07/17 12:04:43 UTC
Avatica and retrying in HA mode
I am investigating HA support for the Go Avatica driver.
Let's assume a scenario where we have multiple Avatica servers behind a
load balancer and do not want to use sticky sessions.
Currently, these are some of the questions I have:
1. A connection with an id "123" opened. A server handles the request
and then fails immediately after. We then call a
PrepareAndExecuteRequest using the connection_id of "123". In this case,
what happens? ExecuteResponse does not appear to have a field telling us
the connection_id is invalid.
2. A connection with an id "456" is opened. A server handles the request
and we call PrepareAndExecuteRequest. This returns a resultset. The
server fails at this point. We then call a FetchRequest to fetch more
rows, but the server has no record of this query. If missing_statement
is true, we recreate the statement on a new connection. What is the
expected corrective action if missing_results is true? I am assuming we
use a SyncResultsRequest to sync the results, and if SyncResultResponse
has missing_statement set to true, because the server failed, we
recreate the statement and retry. Are there any cases where
missing_statement is false but missing_results is true (what causes this
scenario)? Also, how about missing_statement = true and missing_results
= false?
3. I noticed that with ExecuteRequest, FetchRequest and
ExecuteBatchRequest, there is sufficient information in the
statement_handle and other fields to allow Avatica to automatically
recreate the statement or result set on the server if it is missing. Is
there any reason why this is not being automatically handled by Avatica?
Cheers,
Francis
Re: Avatica and retrying in HA mode
Posted by F21 <f2...@gmail.com>.
Hey Josh,
Thanks for the clarification!
Francis
On 20/07/2017 4:30 AM, Josh Elser wrote:
>
>
> On 7/17/17 8:04 AM, F21 wrote:
>> I am investigating HA support for the Go Avatica driver.
>>
>> Let's assume a scenario where we have multiple Avatica servers behind
>> a load balancer and do not want to use sticky sessions.
>>
>> Currently, these are some of the questions I have:
>>
>> 1. A connection with an id "123" opened. A server handles the request
>> and then fails immediately after. We then call a
>> PrepareAndExecuteRequest using the connection_id of "123". In this
>> case, what happens? ExecuteResponse does not appear to have a field
>> telling us the connection_id is invalid.
>
> I don't recall the exact context, but there is information passed back
> to definitively know that either a connection or statement do not
> exist server-side. I can dig into the Java source if you're unable to
> find it :)
>
>> 2. A connection with an id "456" is opened. A server handles the
>> request and we call PrepareAndExecuteRequest. This returns a
>> resultset. The server fails at this point. We then call a
>> FetchRequest to fetch more rows, but the server has no record of this
>> query. If missing_statement is true, we recreate the statement on a
>> new connection. What is the expected corrective action if
>> missing_results is true? I am assuming we use a SyncResultsRequest to
>> sync the results, and if SyncResultResponse has missing_statement set
>> to true, because the server failed, we recreate the statement and
>> retry. Are there any cases where missing_statement is false but
>> missing_results is true (what causes this scenario)? Also, how about
>> missing_statement = true and missing_results = false?
>
> I don't think that missing_statement=false/missing_results=true would
> ever happen as server failure would be the only possible way (the
> avatica server would have crashed and the statement would be lost
> anyways).
>
> A missing statement without missing results may happen (same scenario
> as the first question). I believe it is a true statement that we would
> only be missing results if we were not missing a statement (as the
> results are created by the statement). Similarly, we would never have
> results if we were missing a statement.
>
> If it helps to visualize in code, in JDBC:
>
> Connection conn = getConnection(conn_id); // missing_connection
> Statement stmt = getStatement(conn, stmt_id); // missing_statement
> ResultSet results = getResults(conn, stmt); // missing_results
>
> We never get to the further point if we miss context earlier.
>
>> 3. I noticed that with ExecuteRequest, FetchRequest and
>> ExecuteBatchRequest, there is sufficient information in the
>> statement_handle and other fields to allow Avatica to automatically
>> recreate the statement or result set on the server if it is missing.
>> Is there any reason why this is not being automatically handled by
>> Avatica?
>
> Presently, the server doesn't hold on to that information. For the HA
> case, the server you're currently talking to never saw the creation of
> a connection, statement, results. That's why it has to be client-driven.
>
> We could try to make the Avatica server smart enough to re-create the
> objects when they were automatically expired (after 10mins: default),
> but the common case is that the previous server died and the new
> server just doesn't know.
>
>> Cheers,
>> Francis
>>
Re: Avatica and retrying in HA mode
Posted by Josh Elser <jo...@gmail.com>.
On 7/17/17 8:04 AM, F21 wrote:
> I am investigating HA support for the Go Avatica driver.
>
> Let's assume a scenario where we have multiple Avatica servers behind a
> load balancer and do not want to use sticky sessions.
>
> Currently, these are some of the questions I have:
>
> 1. A connection with an id "123" opened. A server handles the request
> and then fails immediately after. We then call a
> PrepareAndExecuteRequest using the connection_id of "123". In this case,
> what happens? ExecuteResponse does not appear to have a field telling us
> the connection_id is invalid.
I don't recall the exact context, but there is information passed back
to definitively know that either a connection or statement do not exist
server-side. I can dig into the Java source if you're unable to find it :)
> 2. A connection with an id "456" is opened. A server handles the request
> and we call PrepareAndExecuteRequest. This returns a resultset. The
> server fails at this point. We then call a FetchRequest to fetch more
> rows, but the server has no record of this query. If missing_statement
> is true, we recreate the statement on a new connection. What is the
> expected corrective action if missing_results is true? I am assuming we
> use a SyncResultsRequest to sync the results, and if SyncResultResponse
> has missing_statement set to true, because the server failed, we
> recreate the statement and retry. Are there any cases where
> missing_statement is false but missing_results is true (what causes this
> scenario)? Also, how about missing_statement = true and missing_results
> = false?
I don't think that missing_statement=false/missing_results=true would
ever happen as server failure would be the only possible way (the
avatica server would have crashed and the statement would be lost anyways).
A missing statement without missing results may happen (same scenario as
the first question). I believe it is a true statement that we would only
be missing results if we were not missing a statement (as the results
are created by the statement). Similarly, we would never have results if
we were missing a statement.
If it helps to visualize in code, in JDBC:
Connection conn = getConnection(conn_id); // missing_connection
Statement stmt = getStatement(conn, stmt_id); // missing_statement
ResultSet results = getResults(conn, stmt); // missing_results
We never get to the further point if we miss context earlier.
> 3. I noticed that with ExecuteRequest, FetchRequest and
> ExecuteBatchRequest, there is sufficient information in the
> statement_handle and other fields to allow Avatica to automatically
> recreate the statement or result set on the server if it is missing. Is
> there any reason why this is not being automatically handled by Avatica?
Presently, the server doesn't hold on to that information. For the HA
case, the server you're currently talking to never saw the creation of a
connection, statement, results. That's why it has to be client-driven.
We could try to make the Avatica server smart enough to re-create the
objects when they were automatically expired (after 10mins: default),
but the common case is that the previous server died and the new server
just doesn't know.
> Cheers,
> Francis
>