---
title: "Storing Mass Spectrometry Data in SQL Databases"
output:
    BiocStyle::html_document:
        toc_float: true
vignette: >
    %\VignetteIndexEntry{Storing Mass Spectrometry Data in SQL Databases}
    %\VignetteEngine{knitr::rmarkdown}
    %\VignetteEncoding{UTF-8}
    %\VignettePackage{MsBackendSql}
    %\VignetteDepends{MsBackendSql,BiocStyle,RSQLite,msdata,microbenchmark,mzR}
---

```{r style, echo = FALSE, results = 'asis', message=FALSE}
BiocStyle::markdown()
```

**Package**: `r Biocpkg("MsBackendSql")`<br />
**Authors**: `r packageDescription("MsBackendSql")[["Author"]] `<br />
**Compiled**: `r date()`

```{r, echo = FALSE, message = FALSE}
library(MsBackendSql)
knitr::opts_chunk$set(echo = TRUE, message = FALSE)
library(BiocStyle)
```

# Introduction

The `r Biocpkg("Spectra")` Bioconductor package provides a flexible and
expandable infrastructure for Mass Spectrometry (MS) data. The package supports
interchangeable use of different *backends* that provide additional file support
or different ways to store and represent MS data. The
`r Biocpkg("MsBackendSql")` package provides backends to store data from whole
MS experiments in SQL databases. The data in such databases can be easily (and
efficiently) accessed using `Spectra` objects that use the `MsBackendSql` class
as an interface to the data in the database. Such `Spectra` objects have a
minimal memory footprint and hence allow analysis of very large data sets even
on computers with limited hardware capabilities. For certain operations, the
performance of this data representation is superior to that of other low-memory
(*on-disk*) data representations such as `Spectra`'s `MsBackendMzR` backend.
Finally, the `MsBackendSql` supports also remote data access to e.g. a central
database server hosting several large MS data sets.


# Installation

The package can be installed with the `BiocManager` package. To install
`BiocManager` use `install.packages("BiocManager")` and, after that,
`BiocManager::install("MsBackendSql")` to install this package.


# Creating and using `MsBackendSql` SQL databases

`MsBackendSql` SQL databases can be created either by importing (raw) MS data
from MS data files using the `createMsBackendSqlDatabase()` or using the
`backendInitialize()` function by providing in addition to the database
connection also the full MS data to import as a `DataFrame`. In the first
example we use the `createMsBackendSqlDatabase()` function which takes a
connection to an (empty) database and the names of the files from which the data
should be imported as input parameters creates all necessary database tables and
stores the full data into the database. Below we create an empty SQLite database
(in a temporary file) and fill that with MS data from two mzML files (from the
`r Biocpkg("msdata")` package).

```{r, message = FALSE, results = "hide"}
library(RSQLite)

dbfile <- tempfile()
con <- dbConnect(SQLite(), dbfile)

library(MsBackendSql)
fls <- dir(system.file("sciex", package = "msdata"), full.names = TRUE)
createMsBackendSqlDatabase(con, fls)
```

By default the m/z and intensity values are stored as *BLOB* data types in the
database. This has advantages on the performance to extract peaks data from the
database but would for example not allow to filter peaks by m/z values directly
in the database. As an alternative it is also possible to the individual m/z and
intensity values in separate rows of the database table. This *long table
format* results however in considerably larger databases (with potentially
poorer performance). Note also that the code and backend is optimized for
MySQL/MariaDB databases by taking advantage of table partitioning and
specialized table storage options. Any other SQL database server is however also
supported (also portable, self-contained SQLite databases).

The *MsBackendSql* package provides two backends to interact with such
databases: the (default) `MsBackendSql` class and the `MsBackendOfflineSql`,
that inherits all properties and functions from the former, but which does not
store the connection to the database within the object but connects (and
disconnects) to (and from) the database in each function call. This allows to
use the latter also for parallel processing setups or to save/load the object
(e.g. using `save` and `saveRDS`). Thus, for most applications the
`MsBackendOfflineSql` might be used as the preferred backend to SQL databases.

To access the data in the database we create below a `Spectra` object providing
the connection to the database in the constructor call and specifying to use the
`MsBackendSql` as *backend* using the `source` parameter.

```{r}
sps <- Spectra(con, source = MsBackendSql())
sps
```

As an alternative, the `MsBackendOfflineSql` backend could also be used to
interface with MS data in a SQL database. In contrast to the `MsBackendSql`, the
`MsBackendOfflineSql` does not contain an active (open) connection to the
database and hence supports serializing (saving) the object to disk using
e.g. the `save()` function, or parallel processing (if supported by the database
system). Thus, for most use cases the `MsBackendOfflineSql` should be used
instead of the `MsBackendSql`. See further below for more information on the
`MsBackendOfflineSql`.

`Spectra` objects allow also to change the backend to any other backend
(extending `MsBackend`) using the `setBackend()` function. Below we use this
function to first load all data into memory by changing from the `MsBackendSql`
to a `MsBackendMemory`.

```{r}
sps_mem <- setBackend(sps, MsBackendMemory())
sps_mem
```

With this function it is also possible to change from any backend to a
`MsBackendSql` in which case a new database is created and all data from the
originating backend is stored in this database. To change the backend to an
`MsBackendOfflineSql` we need to provide the connection information to the SQL
database as additional parameters. These parameters are the same that need to
be passed to a `dbConnect()` call to establish the connection to the
database. These parameters include the database driver (parameter `drv`), the
database name and eventually the user name, host etc (see `?dbConnect` for more
information). In the simple example below we store the data into a SQLite
database and thus only need to provide the database name, which corresponds
SQLite database file. In our example we store the data into a temporary
file.

```{r, warnings = FALSE}
sps2 <- setBackend(sps_mem, MsBackendOfflineSql(), drv = SQLite(),
                   dbname = tempfile())
sps2
```

Similar to any other `Spectra` object we can retrieve the available *spectra
variables* using the `spectraVariables()` function.

```{r}
spectraVariables(sps)
```

The MS peak data can be accessed using either the `mz()`, `intensity()` or
`peaksData()` functions. Below we extract the peaks matrix of the 5th spectrum
and display the first 6 rows.

```{r}
peaksData(sps)[[5]] |>
head()
```

All data (peaks data or spectra variables) are **always** retrieved on the fly
from the database resulting thus in a minimal memory footprint for the `Spectra`
object.

```{r}
print(object.size(sps), units = "KB")
```

The backend supports also adding additional spectra variables or changing their
values. Below we add 10 seconds to the retention time of each spectrum.

```{r}
sps$rtime <- sps$rtime + 10
```

Such operations do however **not** change the data in the database (which is
always considered read-only) but are cached locally within the backend object
(in memory). The size in memory of the object is thus higher after changing that
spectra variable.

```{r}
print(object.size(sps), units = "KB")
```

Such `$<-` operations can also be used to *cache* spectra variables
(temporarily) in memory which can eventually improve performance. Below we test
the time it takes to extract the MS level from each spectrum from the database,
then cache the MS levels in memory using `$msLevel <-` and test the timing to
extract these cached variable.

```{r}
system.time(msLevel(sps))
sps$msLevel <- msLevel(sps)
system.time(msLevel(sps))
```

We can also use the `reset()` function to *reset* the data to its original state
(this will cause any local spectra variables to be deleted and the backend to be
initialized with the original data in the database).

```{r}
sps <- reset(sps)
```

To use the `MsBackendOfflineSql` backend we need to provide all information
required to connect to the database along with the *database driver* to the
`Spectra` function. Which parameters are required to connect to the database
depends on the SQL database and the used driver. In our example the data is
stored in a SQLite database, hence we use the `SQLite()` database driver and
only need to provide the database name with the `dbname` parameter. For a
MySQL/MariaDB database we would use the `MariaDB()` driver and would have to
provide the database name, user name, password as well as the host name and port
through which the database is accessible.

```{r}
sps_off <- Spectra(dbfile, drv = SQLite(),
                   source = MsBackendOfflineSql())
sps_off
```

This backend provides the exact same functionality than `MsBackendSql` with the
difference that the connection to the database is opened and closed for each
function call. While this leads to a slightly lower performance, it allows to to
serialize the object (i.e. save/load the object to/from disk) and to use it (and
hence the `Spectra` object) also in a parallel processing setup. In contrast,
for the `MsBackendSql` parallel processing is disabled since it is not possible
to share the active backend connection within the object across different
parallel processes.

Below we compare the performance of the two backends. The performance difference
is the result from opening and closing the database connection for each
call. Note that this will also depend on the SQL server that is being used. For
SQLite databases there is almost no overhead.

```{r}
library(microbenchmark)
microbenchmark(msLevel(sps), msLevel(sps_off))
```


# Performance comparison with other backends

The need to retrieve any spectra data on-the-fly from the database will have an
impact on the performance of data access function of `Spectra` objects using the
`MsBackendSql` backends. To evaluate its impact we next compare the performance
of the `MsBackendSql` to other `Spectra` backends, specifically, the
`MsBackendMzR` which is the default backend to read and represent raw MS data,
and the `MsBackendMemory` backend that keeps all MS data in memory (and is thus
not suggested for larger MS experiments). Similar to the `MsBackendMzR`, also
the `MsBackendSql` keeps only a limited amount of data in memory. These
*on-disk* backends need thus to retrieve spectra and MS peaks data on-the-fly
from either the original raw data files (in the case of the `MsBackendMzR`) or
from the SQL database (in the case of the `MsBackendSql`). The in-memory backend
`MsBackendMemory` is supposed to provide the fastest data access since all data
is kept in memory.

Below we thus create `Spectra` objects from the same data but using the
different backends.

```{r}
sps <- Spectra(con, source = MsBackendSql())
sps_mzr <- Spectra(fls, source = MsBackendMzR())
sps_im <- setBackend(sps_mzr, backend = MsBackendMemory())
```

At first we compare the memory footprint of the 3 backends.

```{r}
print(object.size(sps), units = "KB")
print(object.size(sps_mzr), units = "KB")
print(object.size(sps_im), units = "KB")
```

The `MsBackendSql` has the lowest memory footprint of all 3 backends because it
does not keep any data in memory. The `MsBackendMzR` keeps all spectra
variables, except the MS peaks data, in memory and has thus a larger size. The
`MsBackendMemory` keeps all data (including the MS peaks data) in memory and has
thus the largest size in memory.

Next we compare the performance to extract the MS level for each spectrum from
the 4 different `Spectra` objects.

```{r}
library(microbenchmark)
microbenchmark(msLevel(sps),
               msLevel(sps_mzr),
               msLevel(sps_im))
```

Extracting MS levels is thus slowest for the `MsBackendSql`, which is not
surprising because both other backends keep this data in memory while the
`MsBackendSql` needs to retrieve it from the database.

We next compare the performance to access the full peaks data from each
`Spectra` object.

```{r}
microbenchmark(peaksData(sps, BPPARAM = SerialParam()),
               peaksData(sps_mzr, BPPARAM = SerialParam()),
               peaksData(sps_im, BPPARAM = SerialParam()), times = 10)
```

As expected, the `MsBackendMemory` has the fasted access to the full peaks
data. The `MsBackendSql` outperforms however the `MsBackendMzR` providing faster
access to the m/z and intensity values.

Performance can be improved for the `MsBackendMzR` using parallel
processing. Note that the `MsBackendSql` does **not support** parallel
processing and thus parallel processing is (silently) disabled in functions such
as `peaksData()`.

```{r}
m2 <- MulticoreParam(2)
microbenchmark(peaksData(sps, BPPARAM = m2),
               peaksData(sps_mzr, BPPARAM = m2),
               peaksData(sps_im, BPPARAM = m2), times = 10)
```

We next compare the performance of subsetting operations.

```{r}
microbenchmark(filterRt(sps, rt = c(50, 100)),
               filterRt(sps_mzr, rt = c(50, 100)),
               filterRt(sps_im, rt = c(50, 100)))
```

The two *on-disk* backends `MsBackendSql` and `MsBackendMzR` show a comparable
performance for this operation. This filtering does involves access to a spectra
variables (the retention time in this case) which, for the `MsBackendSql` needs
first to be retrieved from the backend. The `MsBackendSql` backend allows
however also to *cache* spectra variables (i.e. they are stored within the
`MsBackendSql` object). Any access to such cached spectra variables can
eventually be faster because no dedicated SQL query is needed.

To evaluate the performance of a *pure* subsetting operation we first define the
indices of 10 random spectra and subset the `Spectra` objects to these.

```{r}
idx <- sample(seq_along(sps), 10)
microbenchmark(sps[idx],
               sps_mzr[idx],
               sps_im[idx])
```

Here the `MsBackendSql` outperforms the other backends because it does not keep
any data in memory and hence does not need to subset these. The two other
backends need to subset the data they keep in memory which is in both cases a
data frame with either a reduced set of spectra variables or the full MS data.

At last we compare also the extraction of the peaks data from the such subset
`Spectra` objects.

```{r}
sps_10 <- sps[idx]
sps_mzr_10 <- sps_mzr[idx]
sps_im_10 <- sps_im[idx]

microbenchmark(peaksData(sps_10),
               peaksData(sps_mzr_10),
               peaksData(sps_im_10),
               times = 10)
```

The `MsBackendSql` outperforms the `MsBackendMzR` while, not unexpectedly, the
`MsBackendMemory` provides fasted access.


## Considerations for database systems/servers

The backends from the *MsBackendSql* package use standard SQL calls to retrieve
MS data from the database and hence any SQL database system (for which an R
package is available) is supported. SQLite-based databases would represent the
easiest and most user friendly solution since no database server administration
and user management is required. Indeed, performance of SQLite is very high,
even for very large data sets. Server-based databases on the other hand have the
advantage to enable a centralized storage and control of MS data (inclusive user
management etc). Also, such server systems would also allow data set or
server-specific configurations to improve performance.

A comparison between a SQLite-based with a MariaDB-based *MsBackendSql* database
for a large data set comprising over 8,000 samples and over 15,000,000 spectra
is available
[here](https://github.com/rformassspectrometry/MsBackendSql/issues/15). In
brief, performance to extract data was comparable and for individual spectra
variables even faster for the SQLite database. Only when more complex SQL
queries were involved (combining several primary keys or data fields) the more
advanced MariaDB database outperformed SQLite.



# Other properties of the `MsBackendSql`

The `MsBackendSql` backend does not support parallel processing since the
database connection can not be shared across the different (parallel)
processes. Thus, all methods on `Spectra` objects that use a `MsBackendSql` will
automatically (and silently) disable parallel processing even if a dedicated
parallel processing setup was passed along with the `BPPARAM` method.

Some functions on `Spectra` objects require to load the MS peak data (i.e., m/z
and intensity values) into memory. For very large data sets (or computers with
limited hardware resources) such function calls can cause out-of-memory
errors. One example is the `lengths()` function that determines the number of
peaks per spectrum by loading the peak matrix first into memory. Such functions
should ideally be called using the `peaksapply()` function with parameter
`chunkSize` (e.g., `peaksapply(sps, lengths, chunkSize = 5000L)`). Instead of
processing the full data set, the data will be first split into chunks of size
`chunkSize` that are stepwise processed. Hence, only data from `chunkSize`
spectra is loaded into memory in one iteration.


# Summary

The `MsBackendSql` provides an MS data representations and storage mode with a
minimal memory footprint (in R) that is still comparably efficient for standard
processing and subsetting operations. This backend is specifically useful for
very large MS data sets, that could even be hosted on remote (MySQL/MariaDB)
servers. A potential use case for this backend could thus be to set up a central
storage place for MS experiments with data analysts connecting remotely to this
server to perform initial data exploration and filtering. After subsetting to a
smaller data set of interest, users could then retrieve/download this data by
changing the backend to e.g. a `MsBackendMemory`, which would result in a
*download* of the full data to the user computer's memory.


# Session information

```{r}
sessionInfo()
```