Users of the SQLite library, particularly application developers, who want to access a SQLite database from different systems connected by a network are often tempted to simply open a database connection by specifying a filename which references a database file somewhere within a network filesystem. ("remote database" here) This "file" is then accessed by means of OS API's which permit the illusion of I/O from/to a local file. The illusion is good but imperfect in important ways.
This simple, "remote database" approach is usually not the best way to use a single SQLite database from multiple systems, (even if it appears to "work"), as it often leads to various kinds of trouble and grief. Because these problems are inevitable with some usages, but not frequent or repeatable, it behooves application developers to not rely on early testing success to decide that their remote database use will work as desired.
This diagram shows components and their linkages for reference in the discussion following:
The issues arise from the properties and utilization of the two data/control channels between the above three blocks.
The "API Call" channel carries less information than the "File I/O" channel. API calls to submit queries or specify data modification normally require substantially fewer bits to be passed back and forth than are transferred to/from the database file to store or find the data. Query result retrieval will normally require much more file traffic than API traffic because the data to be returned is rarely to be found without reading unrequested data.
The API Call channel operates at processor main memory speeds (Giga-words/second), with data often passed by reference (and so not copied.) In contrast, even the fastest File I/O channels are slower. They require the data to be copied, usually over a medium requiring bit-serialization. For spinning magnetic media, transfers await platter rotation and head movement, then are limited by spin velocity.
When the File I/O channel includes a network connection, (in addition to some genuine File I/O at its far end), additional slowness is imposed. Even where raw transfer rate does not limit bandwidth, the traffic must still be packetized and buffered at both ends. Additional layers of I/O handlers add scheduling delays. However, slowed transfers are the least significant issue with network filesystems.
The "API Call" channel is highly reliable, to the extent that error rates are unstated and ignored as negligible. The channel fails only when the system loses power (excepting meteorites, etc.)
The "File I/O" channel, when it directly reaches a local storage device, is also highly reliable. (Spinning storage MTBF exceeds 1 million hours, and NVRAM lasts longer.) Local devices also have a characteristic which is critical for enabling database management software to be designed to ensure ACID behavior: When all process writes to the device have completed, (when POSIX fsync() or Windows FlushFileBuffers() calls return), the filesystem then either has stored the "written" data or will do so before storing any subsequently written data.
When network filesystem apparatus and software layers are interposed between filesystem clients and a filesystem over an actual storage device, significant sources of failure and misbehavior are introduced. While network data transfers are error-checked well, transfer packets do not all reliably arrive at their destination once sent. Some packets are clobbered by other packets and must be resent. Under packet clobbering conditions, repeated retries can impose delays exceeding what is needed for similar data to reach local storage. Some portions of what a client writes can end up stored out of time order relative to other portions written.
Because of the disordering and outright data loss which occur in network filesystem writes, it is critical that sets of file writes can be accurately known to be done before a subsequent set of file writes begins. This assurance is obtained by use of robustly designed and correctly implemented fsync() (or equivalent) OS functions. Unfortunately for some applications, network filesystem sync operation can be less robust than local filesystem sync. Attaining robust sync in the face of network packet transport errors is hard, and safeguards are sometimes relaxed in favor of performance.
The bottom line is that network filesystem sync reliability varies among implementations and installations. The design assumptions upon which it relies may hold more true where an application is tested than where it is relied upon. Rely upon it at your (and your customers') peril. See How To Corrupt Your Database Files.
From the above diagram and discussion, it is obvious that performance (aka "speed") is degraded by insertion of a network link into one of the two channels. Consideration of relative traffic volumes between the API Call channel and the File I/O channel reveals that such insertion will have less performance impact at the API Call channel.
Consideration of reliability impact is easier, with a clearer outcome: Inserting a network link into the API Call channel may also result in call failures at times. But if the Client Application has bothered to use SQL/SQLite transactions properly, such failures will only cause a transaction to fail and be rolled back, without compromising the integrity of the data. In contrast, if the network link is inserted into the File I/O channel, transactions may fail (as for the API Call insertion) but with the additional effect that the remote database is corrupted.
These network unreliability issues can be mitigated, completely or to an acceptable degree, by using SQLite in rollback mode. However, the SQLite library is not tested in across-a-network scenarios, nor is that reasonably possible. Hence, use of a remote database is done at the user's risk.
Generally, if your data is separated from the application by a network, you want to use a client/server database. This is due to the fact that the database engine acts as a bandwidth-reducing filter on the database traffic.
If your data is separated from the application by a network, you want the low-traffic link to be across the network, not the high-traffic link. This means that the database engine needs to be on the same machine as the database itself. Such is the case with a client/server database like PostgreSQL. SQLite is different in that the database engine runs on the same machine as the application, which forces the higher-traffic link to traverse the network in remote database scenarios. That normally results in lower performance.
Network filesystems do not support the ability to do simultaneous reads and writes while at the same time keeping the database consistent. So if you have multiple clients on multiple different machines which need to do simultaneous database reads and writes, you have these choices:
1. Use a client/server database engine. PostgreSQL is an excellent choice. A variation of this is:
2. Host an SQLite database in WAL mode, but do all reads and writes from processes on the same machine that stores the database file. Implement a proxy that runs on the database machine that relays read/write requests from remote machines.
3. Use SQLite in rollback mode. This means you can have multiple simultaneous readers or one writer, but not simultaneous readers and writers.
Application programmers should be cognizant of the possibility that their application's users will elect to use a remote database if they can do so. Unless one of the above choices has been effected, or one at a time, exclusive access is used, a programmer should consider blocking that election unless reliability is of little importance.
Choose the technology that is right for you and your customers. If your data lives on a different machine from your application, then you should consider a client/server database. SQLite is designed for situations where the data and application coexist on the same machine. SQLite can still be made to work in many remote database situations, but a client/server solution will usually work better in that scenario.
*** DRAFT ***