I don't think any realistic implementation keeps information that increases at a rate this implies. Undo based systems have the undo size grow roughly at the amount of changes that aren't visible to everyone, and postgres style in-heap MVCC systems have similar amount of data growth inside the table. Usually the amount of that additional state is more correlated with the age of the oldest observer than with the number of concurrent observers.
> If you allow clients unlimited concurrency, they can execute enough transactions to fill memory and disk with MVCC world-states, causing potential data loss (since the full disk may have prevented the DB engine from writing down a committed world-state.)
I don't think any real-world DB implementation would work that way. The definition of committed in most of such systems include that the necessary data has been written to disk before signalling success for the commit.
This problem also seems largely independent of connection/transaction limits.
I did mean to suggest that it's far harder to limit memory allocation, and thus paging/cache thrashing, when you have an unbounded number of transactions each tracking bounded-size changes†, than when you "only" have a bounded number of transactions each tracking bounded-size† changes. (Seems obvious when phrased that way, right?)
† "bounded-size changes" because most RDBMSes have per-connection memory limits, and can-and-will just crash out a connection [with the semantics of a rollback] if it uses too much working memory.
With limited concurrency, you can do individual queries that stall your own connection; and you can do "maintenance" statements like index rebuilds that lock various resources. But it's very hard to put the DBMS in a state where a DBA can't connect to it in order to issue administrative statements to e.g. kill user connections.
Unlimited concurrency (or bounded concurrency if the bound x the per-connection workmem limit > the DB's RAM size) would allow a badly-configured client to "DDoS the RDBMS with established connections", such that the DBMS could become entirely unresponsive to the network and would need to be non-gracefully terminated.
> The definition of committed in most of such systems include that the necessary data has been written to disk before signalling success for the commit.
True, by definition, yeah. And probably always going to be true in single-instance clusters. In distributed clusters, testing by Jepsen and others has shown that, in most SQL DBMSes with configurable SQL "transaction isolation" levels, using anything other than the highest level configurable can allow for write-loss of locally [but not distributed-ly] "committed" writes, under certain conditions.
Are large numbers of concurrent transactions really an issue in practice though? There may be some per-transaction overhead, but I haven't seen any stern warnings about it, e.g. in the Postgres docs for max_connections.
> Thus, a backend concurrent connection (i.e. transaction) cap of N, translates to O(N) concurrent MVCC "world states" that the DB engine has to keep track of. No matter how efficient the DB engine is at doing that, that's still a lot of state!
I don't quite see what you mean about world state. An abort just rolls back changes made by the aborted transaction, so I would expect SQL engines to store a diff of each updated row. And if that's the case, it seems like a single huge transaction could use as much memory as many small transactions, if the same rows were modified.
There's definitely scalability implications to higher number of established connections in postgres. Essentially computing a "snapshot" (visibility information for queries etc, needs to be computed at least once a transaction) is O(#established-connections). If you get in the high hundreds on a large NUMA machine, and you run tiny fast transactions, you definitely can observe the overhead of that.
>> > Thus, a backend concurrent connection (i.e. transaction) cap of N, translates to O(N) concurrent MVCC "world states" that the DB engine has to keep track of. No matter how efficient the DB engine is at doing that, that's still a lot of state!
There's definitely ways to handle that better than O(N), at least in the average case.
O(N) root-level snapshot objects of the state, I mean. Like O(N) git commits. The minimum size of the persisted MVCC snapshot data, probably just scales with O(log N) of the number of connections, depending on what sort of data structure you're using for copy-on-write'ing your snapshots.
A transaction, in SQL, requires the ability for the client (or the backend) to roll it back at any time before it gets committed—and thus, a transaction in progress requires keeping around at least two "world states" of the DB engine: the one before the transaction began, and the way the world looks after the current statement's modifications—but not those of any other concurrently-executing transaction.
Thus, a backend concurrent connection (i.e. transaction) cap of N, translates to O(N) concurrent MVCC "world states" that the DB engine has to keep track of. No matter how efficient the DB engine is at doing that, that's still a lot of state!
Imagine an RDBMS instance shared between numerous clients built and run by different companies. This RDBMS is locked down in various ways to guarantee each company an SLA on their queries. And, of course, the most important part of that SLA is that a query that is ACKed by the server as having gone through, must never lose data.
If you allow clients unlimited concurrency, they can execute enough transactions to fill memory and disk with MVCC world-states, causing potential data loss (since the full disk may have prevented the DB engine from writing down a committed world-state.)
Thus, ACID-conformant SQL-speaking databases are always going to limit the number of concurrent transactions, and so the number of concurrent connections; and so—at least for SQL-speaking databases designed for instance-addressable rather than cluster-addressable deployment—the responsibility for deciding how much concurrency a client wants/needs is always going to be pushed to the client, so that an enterprise with a DB cluster can set up its less concurrency-intensive clients to self-limit their pool size.
† SQL isn't a wire protocol per se; it's just a syntax and semantics of a formal language, without even a shared encoding. But the standard does specify per-connection and per-transaction state-machines the backend must have, and how these are affected by executing particular statements. So, in this way, SQL is an abstract contract for the semantics that any particular RDBMS's wire protocol must obey. A "meta-protocol", if you will.