The period-change behaviour of 134 RR Lyrae stars in the globular cluster Messier 3 (M3) is investigated on the ∼120 yr time-base of the photometric observations. The mean period-change rates (β≈ 0.01 d Myr−1) of the subsamples of variables exhibiting the most regular behaviour are in good agreement with theoretical expectations based on horizontal branch stellar evolution models. However, a large fraction of variables show period changes that contradict the evolutionary expectations. Among the 134 stars studied, the period-change behaviour of only 54 variables is regular (constant or linearly changing), slight irregularities are superimposed on the regular variations in 23 cases and the remaining 57 stars display irregular period variations. The light curve of ∼50 per cent of the RRab stars is not stable, that is, these variables exhibit Blazhko modulation. The large fraction of variables with peculiar behaviour (showing light-curve modulation and/or irregular O−C variation) indicates that, probably, variables with regular period changes incompatible with their evolutionary stages also could display some kind of instability of the pulsation light curve and/or period, but the available observations have not disclosed it yet. The temporal appearance of the Blazhko effect in some stars, and the 70–90 yr long regular changes preceded or followed by irregular, rapid changes of the pulsation period in some cases, supports this hypothesis.
Accurate Fourier parameters of the light curves of the RRab variables are derived from all the available CCD data. The large, homogeneous sample of stars on the Oosterhoff I sequence enable us to investigate the characteristics of the light curves of the variables showing increasing and decreasing period changes. It is found that, at a given phase-difference value, period-increasing variables have 0.002–0.014 mag smaller amplitudes on average than period-decreasing variables have. Also, at a given period, their phase differences tend to be smaller by 0.03–0.07 rad than the phase differences of variables with decreasing periods. The realness of the detected differences is proven by Monte Carlo simulations.